Model Guidance
Last refreshed: 2026-10-02 (trigger: gemini-4-argon β new-model-detected. THE TRIGGER MODEL IS UNROUTABLE AND THE URGENT FINDING IS A SHUTDOWN DATE NOBODY WAS WATCHING. (1) gemini-4-argon (Google, announced 2026-09-30) is a real Google recovery β AA Intelligence Index v4.3.2 53 (high), #8 of 224, which is +23 over gemini-3.1-pro-preview's 30 and puts Google back in the top three labs β but there is nothing to pin: no published API model id (Google's post names the model and never gives an API string; DataCamp explicitly declined to guess one), not on OpenRouter, Vertex, Gemini CLI, Cursor or Copilot, and access is gated to cyber defenders in the Fairwind Program with no GA date. AA benchmarked it through a single provider (Google AI Studio). Genuine firsts: 1M OUTPUT tokens (up from 64K, via "Long Decode Continuation"), #1 on AutomationBench-AA at 78% (+7 over sonnet-5-5 max), a 15% AA-Omniscience hallucination rate β lowest of any model scoring 45+, and a defensive-cyber posture Google shipped deliberately WITHOUT cyber guardrails to trusted defenders (a third shape after Anthropic's refuse-and-fall-back and OpenAI's Daybreak gating). Vendor numbers: DeepSWE v1.1 77.9, LVBench 91.7 (long-video SOTA), CWE-bench v1 68 (tied 1st), AutomationBench-Zapier 51.3. FOUR TRAPS, all recorded: (a) the $2/$10 is an undated INTRODUCTORY price that reverts to $4/$20 β cost-per-task goes $1.99 β $3.98, ~22% ABOVE GPT-6 Astra's $3.26, so the entire cost advantage is a promotion; (b) 62k output tokens/task vs Astra's 27k for equal measured quality β the same verbosity tax now flagged on opus-5-5, sonnet-5-5 and grok-4.7, and the fourth consecutive launch where it appears; (c) the hallucination headline is NOT an accuracy headline β 15% hallucination sits on 50% AA-Omniscience accuracy vs Astra's 63%, and the-decoder reads the composite knowledge score as 42 vs Astra 43, so it abstains more rather than knowing more (this matters directly to the grok-4.20 honesty seat, which is justified on Omniscience); (d) Google's own claim does not survive AA β TechCrunch repeats Google's "scored significantly higher than GPT-6 Astra and Anthropic's Fable and Opus models" (citing Vals), while AA has it TIED with Astra at 53 and five behind opus-5-5 (58); the-decoder's independent read is "closes the gap but doesn't take a clear lead". Cutoff, effort tiers, SynthID watermark status and ZDR posture are all UNSTATED β recorded null, not inferred. WATCH, zero routing β and the gate is ACCESS, not evidence. (2) THE BIGGER FINDING IS A MISS: gpt-6.1-sol shipped at DevDay 2026-09-29 β ONE DAY BEFORE the 09-29 refresh β and was not caught. It supersedes gpt-6-sol at the IDENTICAL $2/$10 while moving AA v4.3.2 from 48 β 52, and it resets the cost frontier rather than the intelligence one: 52 max @ $0.72/task Β· 51 xhigh @ $0.39 Β· 50 high @ $0.32 Β· 48 medium @ $0.21. Its high tier is cheaper per task than any tier of any other model in AA's top 25, and it reaches ~90% of opus-5-5's index score at ~12% of its per-task cost. 1.1M ctx / 128K out, cutoff Apr 2026, cached input $0.10 (from $1.00), long-context tier >272K rebills at a mild $4/$15 (vs Astra's ~$20/$75 whole-request rebill). On OpenRouter. Caveats: every coding/computer-use number is OpenAI's own harness (DeepSWE +6.4 over gpt-6-sol and matching Astra; AutomationBench +2.2 over opus-5-5 at ~β
cost; OSWorld 2.0 within 2.1 of Astra at ~1/7 cost/task), openai.com returned 403 this run so the vendor announcement is NOT in hand, and the gpt-5.6-luna MRCR 41.3 long-context collapse is still untested across the whole gpt-6 line. Rostered; it retargets the council OpenAI seat for the second time in ten days. (3) A HARD SHUTDOWN DATE, 21 DAYS OUT, ON A LIVE PIN β this is the run's only real deadline. Verified at OpenAI's deprecations page today: gpt-4.1-nano shuts down 2026-10-23, and it is live as BUDGET_MODEL at scripts/photos/extract.py:47. Nothing else on the board has a date this close. Also newly dated: gpt-4o-mini-transcribe removed 2027-02-27 (live at api/src/services/agenda_realtime.py:30). (4) AND THE SAME FETCH CLEARED 11 FALSE POSITIVES: gpt-4o-mini is NOT deprecated at OpenAI β nor is gpt-4.1-mini. The audit classified both as "gpt-4 family (retired)" and tagged 12 refs on that basis; the 09-23 and 09-29 reviews both said "not verified at OpenAI this run β verify before sweeping". Verified today: do not sweep them. That is 12 of the 75 stale refs retired by reading one page, and the carried caveat is now CLOSED. (5) DEPRECATION β the first one this doc has ever had to act on, and it landed the day AFTER the last refresh. claude-sonnet-4-5-20250929 was formally deprecated 2026-09-30, retiring 2026-11-30, replacement claude-sonnet-5-5. Four consecutive reviews carried its floor as a watch item and predicted "nothing can fail before ~late Nov" β that was right almost to the day. Blast radius is NIL: zero live routing pins (grep of scripts/, flows/, .aloop/), the geo.py pin those reviews flagged already reads claude-sonnet-5, and the only two hits are historical price rows at costs.py:40 + cost_report.py:40 that must STAY or past spend becomes unreadable. Record the notice, close the four-review watch, change no code. claude-haiku-4-5's floor is 13 days out and STILL carries no notice β and the sonnet-4-5 pattern says to expect one shortly AFTER the floor passes, not before. claude-haiku-5-5 re-verified absent from the platform model table (still four current models: Fable 5.1 / Opus 5.5 / Sonnet 5.5 / Haiku 4.5) β the null-specs row stands unchanged. (6) ZERO of the five 09-29 recommendations were applied (3 days, no action β contrast 09-21, approved wholesale and applied the same evening). Re-asserted as one batch, not re-litigated: config.json:39 + :90 + research-proposal:41 sonnet-5 β sonnet-5-5, and the costs.py sonnet-5-5 row. No new sonnet-5-5 argument is offered; the 09-29 one stands. (7) AUDIT β aloop_stale=[] IS WRONG A SECOND CONSECUTIVE RUN, for the SAME unfixed reason, plus a new double-count. known_roster still lacks opus-5-5, sonnet-5-5, gpt-6-sol, gpt-6-luna, deepseek-v4.1-flash (and now gpt-6.1-sol, gemini-4-argon), and SUPERSEDED (snapshot_routing.py:77-87) is still a full generation stale β so claude-sonnet-5 remains on the roster and absent from the map, and both live sonnet-5 pins still read CLEAN while being this run's headline recommendation. The 09-29 fix was a data edit to two structures and it did not happen. NEW: .aloop/config.json:71 is double-counted β the sleep_build opus-4-7 pin is correctly held as an aloop mode and simultaneously stale as a hardcoded ref, because DELIBERATE_HOLDS matches mode-name-or-path and .aloop/config.json is neither. A zero is not evidence until the filter is proven able to return something β and a 75 is not evidence until you read what is in it. (8) THE HOLD TRIO IS NOW A PAIR: scripts/reports/generate_html_v2.py NO LONGER EXISTS (only stale .pyc remains). DELIBERATE_HOLDS already reflects this; the routing row below said "three sites" and is corrected. The opus-4-7 reaffirm window opens this month. (9) The 4-vendor research-proposal panel is stale in THREE of four arms, not one. flows/research-proposal/FLOW.yaml:41 reads ["openai/gpt-4.1-mini", "google/gemini-2.5-flash-preview-05-20", "deepseek/deepseek-r1", "anthropic/claude-sonnet-5"] β the 09-29 review proposed bumping only the Anthropic arm and never looked at the other three. Council UNCHANGED on the Claude side (opus-5-5 keeps the seat; argon declined β 53 < 58, and its Omniscience lead is calibration-driven at 50% accuracy, so it does not challenge grok-4.20 either). Est. cost delta β $0/mo (range β$1 to +$3; every swap here is per-token-neutral or cheaper, and this is honestly not a savings run β the one real cost lever, gpt-6.1-sol at $0.32/task, sits on frontier-cheap paths vita does not route to yet). Full report: data/reports/2026-10-02-model-guidance-review.md.) Prior refresh (09-29, claude-sonnet-5-5 + claude-haiku-5-5): ONE OF THE TWO TRIGGER MODELS SHIPPED; THE OTHER IS ANNOUNCED-BUT-UNSHIPPED β and the difference is the whole story of this run. (1) claude-sonnet-5-5 (Anthropic, GA 2026-09-28, claude-sonnet-5-5) is a direct successor to sonnet-5 at IDENTICAL $2/$10 (cache read $0.20, cache write $2.50), 1M ctx / 128K out, cutoff Jun 2026, adaptive thinking with five effort levels defaulting to high, retirement floor 2027-09-28, ZDR-clean, OpenRouter day-one across 5 providers. AA-independent: Intelligence Index v4.3.2 = 56 (max), #3 of 216 β up from sonnet-5's 38 (#58), the largest same-price gen-over-gen jump this doc has recorded; xhigh 52, high 47. Anthropic's table has it BEATING opus-5-5 on Terminal-Bench 4.0 (70.6 vs 66.4) and near-tying it on GDPval-AA v2.1 (1844 vs 1846), AA-Briefcase (1811 vs 1822), OSWorld 2.1 (80.1 vs 81.8) and CursorBench 4.0 (55.5 vs 57.8) β at half the per-token price. THE COST TRAP, AGAIN, AND WORSE THAN OPUS 5.5's: Anthropic says "up to 30% LESS per task" than sonnet-5; AA measures $7.60/task vs sonnet-5's $5.09 β ~49% MORE, and ABOVE opus-5-5's $5.98. AA calls it "notably fast, however very verbose" (410M output tokens in eval, vs sonnet-5's 370M). Both recorded; they measure different things (AA = fixed hard-reasoning at max effort, where a smarter model thinks longer; Anthropic = iteration counts on real customer work β Balyasny 121K tokens/answer vs 497K, Base44 3.6 build iterations vs 7.7). Do not let either number stand alone. Worse for a naive adopt case: on AA's effort frontier opus-5-5 at xhigh scores the same 56 for $3.46, so sonnet-5-5-at-max is strictly dominated by the flagship at a lower tier, and opus-5-5 medium (51 @ $1.34) beats sonnet-5-5 high (47 @ $1.08) on intelligence for a quarter more. The honest lane is speed (137.6 TPS, fastest Sonnet yet) and real-workload token efficiency β NOT AA cost-per-task. New integration notes: higher-risk cyber tasks visibly fall back to sonnet-5 (a different fallback target than opus-5-5's opus-4-8 redirect β first Sonnet with cyber safeguards at all); first Sonnet with anti-distillation preserved-thinking classifiers (same account-creation-date trap flagged for opus-5-5 on 09-23); watermarking is NOT stated on the launch page β recorded UNKNOWN, not inferred from the post-08-02 rule. (2) claude-haiku-5-5 DID NOT SHIP. Anthropic's own Sonnet 5.5 page commits it to "the coming weeks" with no ID, no date, no price, no benchmarks, and the platform model table still lists Haiku 4.5 as the only Haiku. Rostered with null specs β this is the 09-21 phantom lesson handled correctly in its other direction: there IS a primary source, so its existence as a committed release is a fact while its specs are genuinely UNKNOWN. Do not extrapolate from haiku-4-5's $1/$5 β sonnet-5-5 held price exactly while opus-5-5 CUT 20%, so there is no 5.5-generation pricing rule. (3) ROUTING β 3 clean swaps, 2 DELIBERATE HOLDS, and the holds are the finding. sonnet-5 is live at five sites and only three are one-line swaps. scripts/feed/constitutions/extractor.py:34+72 must NOT be blind-swapped: it passes thinking={"type":"disabled"}, which adaptive-thinking Claudes reject with a 400 (fable-5/5.1 both 400; opus-5-5 "thinking can no longer be disabled") β and simply deleting the param re-opens the exact bug the 09-21 pass fixed, where adaptive thinking eats the max_tokens=400 budget and truncates the JSON mid-object. It needs a max_tokens raise AND the param removed, or it holds. scripts/seo/geo.py:44 must NOT be bumped either: those are DataForSEO model_name values, not vendor API ids (the file says so in a comment, and gates changes on GET /v3/ai_optimization/claude/llm_responses/models), and the file deliberately freezes versions for run-over-run comparability. sonnet-5's floor is 2027-06-30 β there is no forcing function, so the 09-21 bump was retirement-driven and this one would not be. (4) AUDIT β aloop_stale=[] IS WRONG THIS RUN, and for a NEW reason. The 09-21 structural fix (roster-driven classification) still holds, but the roster has fallen behind by maintenance lag: known_roster is missing opus-5-5, sonnet-5-5, gpt-6-sol, gpt-6-luna, deepseek-v4.1-flash, and superseded_map is one full generation stale (claude-sonnet-4.6βclaude-sonnet-5 should be βsonnet-5-5; opus-4.8/4.7βopus-5 should be βopus-5-5; gpt-5.5βgpt-5.6-sol should be βgpt-6-sol). Because claude-sonnet-5 is in known_roster and is NOT a key in superseded_map, both live sonnet-5 pilot pins read CLEAN when they are in fact stale. A zero is not evidence until the filter is proven able to return something. (5) AUDIT POSITIVES β two long-carried items CLOSED. The six downstream council flows that this doc had promised to sync since 07-25 are all now on opus-5-5 + gpt-6-sol, as are flows/council/setup_models.py and scripts/council/core.py β all eight council entry points agree for the first time in the doc's history. And the cost-tracking gap is closed: costs.py:27-38 now carries opus-5-5/opus-5/fable-5-1/sonnet-5 with per-model cache multipliers and an explicit "never derive it" comment. Only claude-sonnet-5-5 needs adding (identical rates to sonnet-5). (6) No new deprecations. Every roster Claude reads Active with no notice. But sonnet-4-5's floor passed today (09-29) and haiku-4-5's is 10-15 β both still notice-free, so the 60-day rule pushes any real failure to ~late Nov. haiku-4-5 is now a NON-ISSUE for vita: zero live haiku pins remain (the three were moved to sonnet-5/flash-lite on 09-21). The quieter change: Anthropic's overview page moved sonnet-5, sonnet-4-6 AND sonnet-4-5 into "Legacy (still available)" while the deprecation table still calls them Active β so vita's live balanced default (sonnet-4-6, 26 refs) and its running pilot (sonnet-5) are both off the recommended list. Council UNCHANGED (sonnet-5-5 is not a seat β 56 < opus-5-5's 58, no Omniscience lead, and council volume is too low for the price delta to matter). Est. cost delta β $0 to +$3/mo (all three swaps are per-token identical; direction depends entirely on which verbosity read holds). Full report: data/reports/2026-09-29-model-guidance-review.md.) Prior refresh (09-23, claude-opus-5-5): THE 09-21 "PHANTOM" CALL WAS WRONG. claude-opus-5-5 SHIPPED 2026-09-22 and is now the roster's Frontier-lead, superseding opus-5. Every number in the X-sourced "leaked price sheet" the 09-21 review refused to trust β $4/$20, cache-read $0.20, cache-write $5 β matches platform.claude.com exactly. Withholding the roster row on 09-21 was correct (no primary source existed yet); calling the model non-existent was the overclaim, and it is the second time a dismissed model-status rumor proved real (the first: the Jun-2026 Fable-5 export-control suspension, dismissed 06-23, confirmed 07-05). The standing lesson stands and now has two data points: quarantine unverified rumors, never assert absence from them. (1) claude-opus-5-5 (Anthropic, GA 09-22, claude-opus-5-5) β $4/$20 (20% UNDER opus-5's $5/$25), cache read $0.20 (60% cut from $0.50), cache write $5/$8, batch $2/$10, fast mode $8/$40, 1M ctx / 128K out, cutoff Jun 2026, effort lowβmax (default medium), retirement floor 2027-09-22, OpenRouter day-one (anthropic/claude-opus-5.5). AA-independent: Intelligence Index 58 (max) on v4.3.2 β #1 of 212, a FIVE-point lead over both Fable 5.1 (53) and GPT-6 Astra (53); 54 high / 51 medium; four of five effort tiers on AA's Pareto frontier. It leads six of the ten index components: HLE 61.4 (prev best 59.1), SciCode 66.9 (prev 63.1), GDPval-AA v2.1 1846 Elo (+111 vs Fable 5.1, +138 vs Opus 5), AA-Briefcase v1.1 1822 Elo (+143 vs Fable 5.1), AA-Omniscience, and AutomationBench-AA; behind on CritPt, AA-LCR, GDP.pdf. Anthropic's table: Terminal-Bench 4.0 66.4 (Fable 5.1 55.8 / Opus 5 52.3), CursorBench 4.0 57.8 (51.8 / 46.6), FrontierCode v1.1 54.4 (50.3 / 48.0), OSWorld 2.0 81.8 (80.7 / 74.0), Chartography 89.0, Terminal-Bench-Science 58.7, HLE-with-tools 67.7. THE COST TRAP β cheaper per TOKEN is NOT cheaper per TASK: AA measures cost-per-task at $5.98 vs Opus 5's $5.86 β flat, because 5.5 burns ~1.6Γ the output tokens (~119k/task; 260M across AA's eval vs an 88M median β AA calls it "very verbose"). AA's own words: "maintains cost parity with Opus 5 despite 1.6x higher token usage." The 20% cut buys a large intelligence gain at the SAME agentic spend; the real savings land on cache-heavy and short-output paths, where the 60% cache-read cut dominates. Any estimate that multiplies opus-5 volume by 0.8 is wrong for agentic work. BENCHMARK DISAGREEMENT (both recorded): Terminal-Bench 4.0 β Anthropic says 66.4 and puts Astra at 57.9; AA's independent run says 59.6 and calls it a TIE with Astra. Different harnesses; AA additionally ran AutomationBench WITHOUT fallback models, so safeguard interventions counted as outright failures. Prefer AA cross-vendor, Anthropic gen-over-gen. ZDR available (keeps opus-5's data-posture advantage over Fable 5.1). NEW: it carries the EU-AI-Act invisible watermark β opus-5 was the last un-watermarked frontier Claude, so anything published AS bio-zack now has the same consideration on Opus as on Fable. Cyber: redirects most cyber tasks to opus-4-8 β that fallback row stays load-bearing; NOT an authorized-security-testing option (Fable 5.1 keeps the discover-not-exploit GA path). Thinking can no longer be disabled. New integration trap: a "preserved thinking anti-distillation safeguard" is active for API accounts created after 2026-08-31 β verify which vita account that hits before assuming opus-5 thinking-block behavior carries. Tool-use system prompt overhead dropped to a flat 286 tokens (opus-5 charged 406 for any/tool). Subscription side: 5-hour caps raised/removed on Pro/Max/Team/seat-Enterprise + a rate-limit reset β relevant to bio-zack's 09-22 complaint about burning banked resets on Astra. (2) A SAME-DAY FRONTIER PRICE WAR β GPT-6 Sol and GPT-6 Luna also shipped 09-22, 19 days after Astra, cutting OpenAI API prices 50%+. gpt-6-sol $2/$10 (from Astra-era $4/$20; cached $0.20; 872K ctx per AA; 125 TPS; AA v4.3.2 Intelligence Index 48 (max), cost-per-task $0.13 low β $1.06 max) β that is ten points BELOW Opus 5.5's 58 but at roughly a SIXTH of its per-task cost. Flag the vendor-vs-independent gap: launch coverage repeats OpenAI's claim that Sol "beats Opus 5 at 9% of the cost per task" on AutomationBench, but AA's composite has Sol 48 < Opus 5 51 < Opus 5.5 58 β a single-benchmark win is not a composite win. gpt-6-luna $0.10/$0.50 (from $0.20/$1.20), ~1.05M ctx / 128K out β roughly a 10Γ cut against the rostered gpt-5.6-luna ($1/$6) and cheaper on output than the budget-fast pin gemini-3.1-flash-lite (~$0.10/$0.40). Both rostered; Luna's gpt-5.6-era MRCR 41.3 long-context collapse must be re-tested, not assumed inherited. This makes the standing "council OpenAI seat gpt-5.5 β gpt-5.6-sol" proposal obsolete β retarget it to gpt-6-sol (see rec 4). (3) DeepSeek V4.1-Flash (GA 09-10, 552B MoE / 8B-16B active, Causal Encoder-Decoder, native vision, 1M ctx, 384K max out) rostered WATCH β its pricing is TIME-OF-DAY dependent, a first for the roster: off-peak $0.15 in / $0.60 out ($0.003 cache-hit), peak DOUBLE, peak window Mon-Fri 01:00-04:00 + 06:00-10:00 UTC. That is 2.1-4.3Γ the rostered v4-flash's flat $0.28 output depending on the hour, and vita's schedulers run inside those windows β not a straight cheaper successor; model the split before any routing. (4) NO new deprecations, but opus-5 quietly moved to "Legacy (still available)" on the models-overview page while the deprecation table still calls it Active with a 2027-07-24 floor β both recorded; practical read is no notice, no risk, but it has stopped being a recommended target. sonnet-4-5 (floor 09-29, six days out) and haiku-4-5 (10-15) still carry no deprecation notice, so the 60-day rule pushes any real failure to ~late November. Sonnet 5.5 + Haiku 5.5 are signalled "within weeks" β which argues for letting the running sonnet-5 pilot ride rather than widening it, and finally gives the haiku-4-5 succession question a real answer. (5) ROUTING β this is the first review in the series where the new model is BOTH the independent #1 AND cheaper per token than the incumbent live default, so unlike the WATCH-only postures given to Astra/Fable-5.1/qwen/glm-5.3, there IS an adopt rationale. 6 proposals: council Claude seat + synthesis opus-5 β opus-5.5; sweep the 6 downstream council flows still pinning opus-4-7/4-8 β two generations stale, promised in this doc since 07-25 and never applied; retarget the OpenAI seat to gpt-6-sol; re-examine the $20/$60 grok-4.20 honesty seat now that a model has finally cleared the AA-Omniscience gate; add opus-5.5/gpt-6-sol/gpt-6-luna cost rows with a per-model cache-multiplier column (opus-5.5's 0.05Γ is a THIRD distinct shape after standard 0.1Γ and Fable's 0.025Γ β a flat cache column mis-costs it); doc/roster update. DELIBERATE HOLD RESPECTED: the opus-4-7 trio (sleep_build, cadence/daemon.py:258, generate_html_v2.py:54-55) NOT re-surfaced β the ~Oct-2026 quarterly reaffirm window opens next month, and opus-5.5 is simply the new eventual target there. Est. cost delta β β$4 to β$10/mo (per-token cuts on low-volume council paths, partially offset by opus-5.5's 1.6Γ verbosity). Full report: data/reports/2026-09-23-model-guidance-review.md.) Prior refresh (09-21, claude-opus-5-5 rumor): the trigger was called a PHANTOM β that call was WRONG, see above. At the time: claude-opus-5-5 did not exist at any primary source β verified absent at the platform.claude.com model table (latest Opus = claude-opus-5), the deprecation table, anthropic.com/news (nothing since the 9-01 Fable 5.1 post), openrouter.ai/anthropic and AA's Anthropic page. The detection came from data/discoveries/2026-09-21.json, which already labeled it an UNVERIFIED X rumor (one account, codename "claude-wafer-eap", claimed release Tue 9-22); a "leaked" price sheet ($4/$20, cache-read $0.20, cache-write $5 β exactly 20% off Opus 5) and a demo video (retracted as recycled 9-18 YouTube footage) arrived from three accounts within three hours. NOT rostered; re-check 9-23 β if it ships, re-run from the real announcement, never from the leaked sheet. What the verification pass found instead: (1) Grok 4.7 (SpaceXAI, GA 09-21, x-ai/grok-4.7) β grok-4.6's successor at the SAME $2/$6 and 500K ctx; AA-independent Intelligence Index 46 on the NEW v4.3.2 scale (#16/202, $3.74/task, 38.8 TPS, AA flags it "notably slow and very verbose" β 240M output tokens in eval); vendor table beats Sol on CursorBench 4.0 (46.3 vs 41.7), EEBench (64.0 vs 39.4) and Terminal-Bench 4.0 (38.0 vs 37.3), trails Fable 5.1 on most rows β rostered as the Frontier-agent-cheap successor, grok-4.6 superseded, ZERO live routing (grok-4.6 was never live-pinned either). (2) Sonnet 5's $2/$10 "intro" price is now PERMANENT β the pricing page states the scheduled 9-01 rise to $3/$15 "will not occur"; the roster row + ΒΆΒΆ footnote were wrong and are corrected. Per-token it is now 33% under sonnet-4-6, the official ~30% tokenizer inflation roughly cancels β cost-neutral-to-slightly-cheaper, so the pilot's cost argument flipped from "gets worse Sep 1" to "stays flat". (3) Meta Muse Spark 1.3 (GA 09-02, $1.25/$4.25 search-snippet pricing, 1M ctx, text+image+video) is 48 on AA v4.3.2 β ABOVE GPT-5.6 Sol (47) and Grok 4.7 (46) at the lowest cost-per-task in the top 12 ($1.60); rostered WATCH (pricing not model-card-confirmed; a $0.10/$0.20 "contributor" tier TRAINS ON YOUR DATA β never route vita there). (4) AA RESCALED AGAIN β Index v4.3 (09-07) / v4.3.2: Terminal-Bench 2.1β4.0, ΟΒ³-Banking β AutomationBench-AA (Zapier, 657 private tasks), private test sets 45% of weight. Every Intelligence-Index number in this doc's v4.1.1 blocks is on a retired scale. New scale: Fable 5.1 max 53 = GPT-6 Astra max 53 Β· Opus 5 max 51 Β· Fable 5.1 high 51 Β· Muse Spark 1.3 48 Β· Opus 5 high 48 Β· GPT-5.6 Sol max 47 Β· Grok 4.7 xhigh 46 β so the 09-04 "Astra 5 below Fable" framing no longer holds (a TIE, and Astra is ~2.3Γ cheaper per task: $3.26 vs $7.63). Cost-per-task column: Fable 5.1 max $7.63 / xhigh $5.98, Opus 5 max $5.86, Astra max $3.26, Sol max $1.99, Muse 1.3 $1.60. (5) Deprecation watch (first real one in months): no current-roster model has an EOL, but claude-sonnet-4-5's retirement floor is 9-29 and claude-haiku-4-5's is 10-15 (60-day notice rule β nothing can fail before ~late Nov, but a notice can now arrive any day). That makes the long-carried scripts/seo/geo.py:41 sonnet-4-5 pin TIME-SENSITIVE, and surfaces three live haiku-4-5 pins (scripts/feed/constitutions/extractor.py:34, scripts/media/bookmarks.py:148, flows/research-proposal/FLOW.yaml:41) with no Haiku 5 announced. (6) Council drift CONFIRMED (closes the 09-04 open item): flows/council/setup_models.py:5-8 still pins opus-4-7 / gpt-5.5 / gemini-3.1-pro-preview / grok-4.20 while scripts/council/core.py:22-26 pins opus-5 / gpt-5.6-sol / gemini / grok-4.5 / kimi-k3 β two council entry points, two rosters. Routing: ZERO live pin moves this run; 7 proposals (rumor-hold, geo.py, setup_models.py sync, sonnet-5 pilot re-decision, grok-4.5β4.7 menu refs, haiku inventory, the 11th-flag frozen snapshot_routing.py:27 roster). DELIBERATE HOLD RESPECTED: opus-4-7 trio NOT re-surfaced (07-18 rule; the ~Oct-2026 reaffirm window opens next month). Est. cost delta β $0 to β$2/mo (sonnet-5 pilot on 2 modes is the only spend change). Full report: data/reports/2026-09-21-model-guidance-review.md.) Prior refresh (09-04, gpt-6-astra): GPT-6 Astra (OpenAI, GA 09-03, gpt-6-astra) is OpenAI's first 6th-gen model, taking GPT-5.6 Sol's top-of-lineup slot β added as a WATCH roster row + β
footnote + a SCORED domain mini-block (it has AA-independent numbers, unlike the WATCH-only qwen/glm-5.3/gemini-3.7-flash entries), zero live routing. $10/$50 (cached input $1, cache-write $12.50; prompts >272K tokens rebill the WHOLE request to ~$20/$75), 1.05M ctx / 128K out, cutoff 2026-04-30, effort lowβmax (no none), NOT on OpenRouter (OpenAI-API-only at launch; first to the Daybreak cyber cohort). AA-independent: Intelligence Index (v4.1.1/"2026" scale) = 61 (max) β a THREE-WAY TIE with GPT-5.6 Sol (max) and Grok-4.6 (high), and 5 points BELOW Claude Fable 5.1 (66/65.7); AA Coding Agent Index 67 β Opus 5. Independent coverage (officechai/trendingtopics) frames the general-intelligence result as underwhelming; AA also logs a ~80-Elo GDPval-AA-v2 DROP vs Sol and a still-high 51% hallucination rate at max effort. Where it genuinely steps up (all three are OpenAI/vendor-harness, not AA-independent): COMPUTER-USE (OSWorld 2.0 72.6% at ~40min/task vs Sol 65.7% @ ~75min β new SOTA), SWE (Terminal-Bench 4.0 57.7% > Fable 5.1 55.8% > Sol 37.3%), and CYBER (FIRST OpenAI model rated CRITICAL under the Preparedness Framework β also HIGH bio/chem; ExploitBench 100% without safeguards vs Sol 78.5%). Two flags: (1) the "Astra 61 = Sol" tie is AA's MAX-effort read β do NOT overwrite this doc's v4.1.1 baseline where Sol reads 58.9 / Fable 5.1 65.7 (effort-tier ambiguity, both recorded); (2) the cyber headline is a TRAP for the security-testing lane β exploit-CREATION ships GATED behind Daybreak and is BLOCKED in the production API (TheHackerNews), so a normal API caller still gets a refuse, exactly like the rest of the roster minus Fable 5.1's new discover-not-exploit GA path β Astra does NOT become vita's authorized-security-testing option. Routing: WATCH, zero live pin moves β $10/$50 is 2Γ Opus 5's $5/$25 and 2Γ Sol's $5/$30 for equal-or-lower measured intelligence; no adopt rationale. Council: evaluated β DECLINED (no AA-Omniscience component, expensive, 51% hallucination β grok-4.20 keeps the honesty seat). The one differentiated future-watch lane is computer-use: vita has no OSWorld-style screen-driving path today, but if one is ever added Astra is the current independent leader β pull an AA/independent computer-use number + Astra's ZDR/watermark/vision posture next refresh before any routing. Audit hygiene (10th consecutive flag): aloop_stale=[] is the frozen snapshot_routing.py:27 CURRENT_ROSTER BLIND SPOT, not clean β the audit's own known_roster still lists opus-4-7/glm-5.1/kimi-k2.6/gpt-5.5, missing ~17 live models incl opus-5/fable-5-1/grok-4.6/astra, so the gauge literally cannot see a stale pin. One genuine live stale ref: scripts/seo/geo.py:41 claude-sonnet-4-5βclaude-sonnet-4-6 (carried from 09-03, still unfixed). Audit POSITIVE: scripts/council/core.py:22/33 now correctly pin claude-opus-5 + :23 gpt-5.6-sol β the 07-25 opus-4-7 drift is RESOLVED in core.py (open: confirm flows/council/setup_models.py matches). Est. cost delta β $0/mo (WATCH = no routing; the reconciliations are price-neutral). Full report: data/reports/2026-09-04-model-guidance-review.md.) Prior refresh (09-03, claude-fable-5-1 + claude-mythos-5-1): Claude Fable 5.1 (Anthropic, GA 09-01, claude-fable-5-1) supersedes fable-5 at the SAME per-token $10/$50 but ~25% cheaper EFFECTIVE (up to ~45% on agentic coding) via token-efficiency + a 75% cache-read cut ($1.00 β $0.25); batch $5/$25; 1M ctx / 128K out, adaptive-thinking-always-on. New #1 on the AA Intelligence Index (65.7 on AA's re-scaled v4.1.1/"2026" index β above Opus 5 63.0 and Fable 5 62.1); Terminal-Bench-Science 24.7β52.6 (Opus 5 29.0, Sol 22.4), GDPval-AA v2 1723β1853 (Opus 5 1824), OSWorld 36.1β41.7, AutomationBench 17.1β31.4, CursorBench 70.5β73.4. Three genuinely-new things beyond the benchmark bump: (1) CYBER β strongest cyber capability of any Anthropic model, ~60% fewer Claude Code cyber-safeguard interventions, and for the FIRST time it can discover software vulnerabilities (still refuses exploit development) β the first GA, generally-reachable option for AUTHORIZED/defensive security-testing where the whole roster (opus-5/opus-4-8/fable-5) refused (glm-5.3's cyber was vendor-only + unweighted, Mythos 5.1 is trusted-access-only); (2) invisible watermark β fable-5.1 + mythos-5.1 are the first Claude models to embed an EU-AI-Act text/file watermark, detectable only via Anthropic's detection API β a new consideration for anything vita publishes AS bio-zack (X, email, PR text, reports); (3) the AA index re-scaled again to v4.1.1, which RESOLVES the 08-13 open flag (Opus-5 now reads 63.0 β AA did rescale; the "61=Sol" parity was scale drift) and drops GPT-5.6 Sol (58.9) BELOW Fable 5 + Opus 5. Mythos 5.1 = same model, cyber/bio safeguards, trusted-access US-orgs-only β NOT vita-adoptable (same posture as Mythos 5). ZERO live routing pins move β fable-5 was never live-routed (niche escalation only), and fable-5-1 STAYS that niche escalation: opus-5 remains the cheaper ($5/$25 vs $10/$50), ZDR-clean, un-watermarked default AND the better council-synthesis pick. Council UNCHANGED (fable-5-1 has no AA-Omniscience component; the standing opus-4-8βopus-5 seat-swap + gpt-5.5βsol proposals still stand, bio-zack's call). The biggest actual finding is a HYGIENE lie β 9th consecutive flag: the audit's known_roster STILL lacks opus-4-8/opus-5/fable-5(-1)/glm-5.2-5.3/grok-4.5-4.6/gpt-5.6-/sonnet-5/kimi-k2.7-k3/qwen/gemini-3.7-3.8-flash β proof the frozen snapshot_routing.py:27 CURRENT_ROSTER never updated β so aloop_stale reads [] because the gauge is BLIND, not because routing is clean (it literally cannot see the sleep_buildβopus-4-7 hold-pin or any superseded pin). New leaderboard-only sightings, all WATCH/zero-routing (no announcement/pricing fetched this run): Gemini 3.8 Flash (AA 58.7), GLM-5.3-Flash (57.5), Meta Muse Spark 1.2 (56.8 β first Meta appearance near the top), Qwen3.8-Flash-Next (55.8); glm-5.3 now finally has a published AA index (59.5). One genuine live stale ref: scripts/seo/geo.py:41 claude-sonnet-4-5 β bump to claude-sonnet-4-6. Est. cost delta β $0/mo (no live pin moves; the cache cut only makes any future fable-escalation cheaper). Full report: data/reports/2026-09-03-model-guidance-review.md.) Prior refresh (08-14, glm-5.3 + gemini-3.7-flash): Two cheap/mid-tier drops within a day of the grok-4.6 refresh, both rostered WATCH-only (no independent AA numbers β zero routing changes). GLM-5.3 (Z.ai/Zhipu, GA 08-14) is a post-training-only refresh of glm-5.2 β SAME 743B base, all gains from extended post-training β that Z.ai calls the strongest open-weights CODING model (+50% over 5.2; DeepSWE 66.9, AutomationBench 48.2, GDPval-AA v2 1769 Elo, all vendor-harness). The genuinely new signal is an EMERGENT CYBER capability β CyberGym 84.5% edges Mythos 5 + GPT-5.6 Sol, and it found 2,436 real vulns across 269 projects (some ~40yr-old); gaps remain vs closed frontier on deep exploitation (ExploitBench). That matters to vita: every Anthropic frontier model refuses offensive-security, so an open-weight, self-hostable, cyber-strong model is the first roster-adjacent option for AUTHORIZED security-testing paths β tracked as a future pilot, gated on weights landing + an independent CyberGym/ExploitBench reproduction. But per-token pricing is UNPUBLISHED (only GLM Coding Plan subs: Lite $12.6 / Pro $56 / Max $117.6 per mo), open weights + license are ~2wk out (glm-5.2 was MIT; do not assume it carries), context unstated (5.2 = 1M usable, likely same, unconfirmed), not on OpenRouter, no AA Intelligence Index β glm-5.2 STAYS the strong-cheap-code pin (no swap). Gemini 3.7 Flash (Google, GA 08-13) is a cheap agentic/coding/document workhorse β big gen-over-gen jumps vs 3.6-Flash (FrontierCode 1.1 34.4β43.6, GDP.pdf 22β34, AutomationBench 17β30.4, all vendor), 1M ctx / 65K max out, multimodal, OpenRouter day-one; Google claims it tops Claude on business workflows (vendor). vita NEVER rostered 3.5/3.6 Flash β the budget-fast pin is still gemini-3.1-flash-lite (~$0.10/$0.40) β so 3.7-Flash is a genuinely NEW mid-tier between flash-lite and gemini-3.1-pro ($2/$12), a candidate cheap-agentic challenger, NOT a flash-lite replacement. PRICING DISAGREEMENT (flagged, both recorded, verify at the model card before any volume routing): news/vendor cite intro $0.75/$3.75 through 2026-12-31 β standard $1.50/$7.50 on 2027-01-01, but OpenRouter's listing surfaced $0.375/$1.875 (exactly half the intro β may be a batch/further-discount rate). No independent AA numbers for either model; both WATCH, zero routing. Council UNCHANGED (neither has an AA-Omniscience component β not a honesty seat; grok-4.20 keeps it). Est. cost delta β $0/mo. glm-5.2 now SUPERSEDED by glm-5.3 (no formal EOL); no current-roster model has an EOL announcement. Carried forward (bio-zack's call, NOT auto-applied): the frozen snapshot_routing.py:27 CURRENT_ROSTER (still opus-4-7, now missing ~15 models incl opus-5/grok-4.6/glm-5.3/gemini-3.7-flash β the "0 aloop stale" gauge still lies by omission, 8th review flagged), the cost-tracking gap (now ~15 models / 11 reviews overdue), the reflect.md:348 placeholder swap, and the council swaps (opus-4-8βopus-5 Claude seat + core.py:22 opus-4-7 drift; gpt-5.5βgpt-5.6-sol gated on METR). DELIBERATE HOLD RESPECTED: the opus-4-7 trio (sleep_build, cadence/daemon.py:258, generate_html_v2.py:54-55) NOT re-surfaced (07-18 self-terminating rule; reaffirm ~Oct 2026). Prior refresh (08-13, grok-4.6): Grok 4.6 (SpaceXAI, GA 08-12) rostered as the successor to grok-4.5's Frontier-agent-cheap tier β SAME $2/$6 base ($4/$12 above a 200K-token prompt, whole request rebilled), SAME 500K ctx, Feb-1-2026 cutoff, OpenRouter/Cursor/Grok-Build day-one. A strict upgrade over grok-4.5: +5 AA Intelligence Index β 61 (SpaceXAI + AA both frame it as matching GPT-5.6 Sol / "Sol Max"), DeepSWE v1.1 54β65.9, APEX-Agents 47.1β57.5, CursorBench v3.2 69.9, FrontierCode v1.1 61.3. AA-verified independent read (direct-fetched the AA analysis article): GDPval-AA v2 1753 Elo β behind ONLY Claude Opus 5, overlapping CIs with Fable-5 + Qwen3.8-Max; AA-Briefcase 1577 Elo (Fable-5-tier, behind the Opus 5 family); cost-per-task $0.84 (ties Kimi K3 at slightly higher intelligence); striking efficiency profile (~53 turns / ~0.5B input tokens vs Opus 5's ~103 / ~2.0B β comparable results in half the turns, a quarter of the input). The catch: its lane is AGENTIC knowledge-work + cost-efficiency, NOT raw coding β DeepSWE 65.9 still TRAILS GPT-5.6 Sol Max (~73), and every SWE-style number is xAI's own harness (no independent SWE-bench Pro/Verified reproduction). Effectively SUPERSEDES grok-4.5 (no formal 4.5 EOL announced). AA-INDEX DISCREPANCY FLAGGED: the AA analysis article implies Opus-5 = 63 while this doc's v4.1 baseline has Opus-5 61 / Sol 59 β so either AA rescaled the index again or the "61 = Sol" claim is against Sol's MAX-effort tier. I did NOT directly fetch the AA grok-4.6 model page this run (numbers are from the AA analysis article + vendor + search snippets), so BOTH readings are recorded β do NOT re-pull the reasoning column or trust the parity claim until a direct AA model-page fetch reconciles it (open question). Still NO published AA-Omniscience component β council UNCHANGED (grok-4.20 keeps the honesty seat); vision/tool-use modality NOT detailed in the announcement (do not route OCR/PDF here until confirmed β same open caveat as grok-4.5). Provider = SpaceXAI (xAI rebrand); verify XAI_API_KEY/base-URL path unchanged. SCORED in a domain mini-block (AA-verified benchmarks, unlike the WATCH-only qwen entry). Routing: grok-4.6 refreshes grok-4.5's still-unrun bounded-coding pilot proposal (cost-per-task challenger on flows/code-review/synthesize); ZERO live routing swap β est. cost delta ~$0. Carried forward (no new info): cost-tracking gap (now 14 models / 10 reviews overdue), frozen snapshot_routing.py:27 CURRENT_ROSTER (the "0 aloop stale" lying gauge), reflect.md:348 placeholder swap; opus-4-7-trio DELIBERATE HOLD respected, not re-surfaced.) Prior refresh (08-05, qwen3.8-max): Qwen3.8-Max (Alibaba/Qwen, GA 08-03, previewed 07-19 at WAIC Shanghai) rostered as a WATCH-tier open-frontier-multimodal entrant β the Qwen family's FIRST appearance on the roster. 2.4T sparse-MoE (~95B active), 1M ctx (max output 131K), native text+image+video, $2/$6 (grok-4.5-equal; one aggregator outlier lists $2.87/$7.12 list / $2.30/$5.69 discounted β flagged), cache-hit implicit $0.25 / explicit-read $0.17, OpenRouter day-one (single provider), open weights promised ~week of Aug-10 (first Max-class Qwen ever open-sourced; license UNSPECIFIED at launch β do NOT assume prior Qwen Apache/Modified-MIT carries). Independent signal is mixed against the vendor hype: AA Intelligence Index 53 (search-snippet β AA model page 404'd on direct fetch this run, NOT model-card-confirmed; presumed v4.1) slots mid-frontier, BELOW open-weight peer Kimi K3 (57) and Terra (55), above Luna (51). Alibaba's own table (no independent SWE-bench reproduction pulled): Terminal-Bench 2.1 86.6 (vs Opus 4.8 84.6, Fable-5 84.6, Sol 88.8), SWE-bench Pro 67.7 (vs Fable-5 80), FrontierSWE 73.5, GPQA-D 92.6, PaperBench 93.0, IFBench 82.8. #5 Text Arena / #2 Vision Arena β the one differentiated lane is cheap frontier MULTIMODAL (native video at $2/$6), a credible future gemini-3.1-pro challenger IF AA vision/OCR numbers materialize. China data-jurisdiction / state-law risk (coverage flags enterprise concern) β non-sensitive paths only, mitigated by self-host once weights land. Verbose (150M tokens in AA eval) + slow. NOT SCORED in the 1-10 grid (AA-53 is search-snippet-only + all coding/vision numbers are vendor β scoring now = fabrication, same discipline as the 07-05 entrants). ZERO routing changes; three gates before any pilot: (1) an independent SWE-bench Pro/Verified or common-harness coding number, (2) weights + license land + a second OpenRouter provider stabilizes capacity, (3) an AA vision/OCR number for the multimodal lane. Prior refresh (07-25, opus-5): Claude Opus 5 (Anthropic, GA 07-24) rostered as the new Frontier-agent leader β supersedes opus-4-8 at the SAME $5/$25 (price-neutral strict upgrade). AA Intelligence Index v4.1 = 61 (max), narrowly #1 of all models (Fable-5 60 / Sol 59 / K3 57); joint-#1 AA Coding Agent Index (Claude Code @ xhigh); record GDPval-AA v2 1861 Elo + AA-Briefcase 1720 Elo (+146 vs Fable-5); beats Fable-5 on OSWorld-2.0 computer-use at ~1/3 cost; ~26% lower cost-per-task than Fable-5; 1M ctx (secondary-source β official page stated price only, confirm at model card). ZDR-CLEAN (no data-retention requirement, unlike Fable-5); cyber classifiers fire ~85% less than Fable-5 but still refuse offensive-security work β opus-4-8 is the documented fallback (wire it); behind Mythos-5 (Glasswing-only) on cyber. Reshapes the roster top: opus-5 is the default frontier pick AND cheaper-per-task than Fable-5 β Fable-5 narrows to a niche (verbose deliberate style; NO independent SWE-bench-Pro opus-5-vs-Fable-5 number pulled β open). Proposals (NOT auto-applied): council Claude seat opus-4-8βopus-5 (SUPERSEDES the Fable-5-synthesis pilot β same-or-better quality at half cost, ZDR-clean, no availability-fallback headache); name opus-5 the coding-agent default; add opus-5 row to costs.py. DELIBERATE HOLD RESPECTED: the opus-4-7 trio (sleep_build aloop, cadence/daemon.py:258, generate_html_v2.py:54-55) stays held per the 07-18 self-terminating rule β opus-5 is simply the new eventual target at the ~Oct-2026 reaffirm; NOT re-surfaced this run. Also flagged: audit shows scripts/council/core.py:22 still pins opus-4-7 (doc claimed opus-4-8 β drift) β fold to opus-5 in the seat swap. Prior refresh (07-18, kimi-k3): Kimi K3 (Moonshot, GA 07-16, 2.8T sparse MoE β largest open-weight model ever announced) rostered as the first open-weight(-promised) model in the frontier tier: AA Intelligence Index v4.1 = 57 (#3 per launch coverage / #4 per AA's own page β discrepancy recorded; behind Fable-5 60 and Sol 59, ahead of Terra 55), $0.94/AA-task, 1M ctx, text+image in, cache-hit $0.30. Pricing regime change: $3/$15 = 4Γ k2.6's input β K3 is premium-band (competes with Terra / grok-4.5 / Sonnet), NOT a k2.6 successor; k2.6 RETAINED on all strong-cheap synthesis sites. Arena.ai frontend-code #1 ahead of Fable-5 (independent); coding/agentic claims (Terminal-Bench 2.1 88.3 KimiCode-harness, FrontierSWE 81.2, SWE Marathon 42.0) are vendor-only β same re-confirm-at-AA posture as glm-5.2. WATCH status, zero routing changes: weights + license land ~Jul-27, OpenRouter upstream capacity-limited (frequent 429s), single 'max' reasoning-effort with real thinking-token overhead. opus-4-7β4-8: DELIBERATE HOLD executed this review per the 07-10 self-terminating rule (6 re-surfaces, declined by inaction; reaffirm ~Oct 2026 β the system stops asking). Prior refresh (07-10): gpt-5.6 Sol/Terra/Luna rostered + scored; AA index rescaled to v4.1; council seat swap gpt-5.5βsol proposed, still pending bio-zack; METR reward-hacking caveat on Sol still unconfirmed at primary.)
Auto-maintained by flows/model-guidance-review (run via stepwise run flows/model-guidance-review --name model-review --wait). See the footer for methodology.*
This doc is the single source of truth for which LLM to use for which task across vita, vita's stepwise flows, gumball, and cadence. It's refreshed whenever a new model drops; see Β§ Methodology.
TL;DR β approved roster (June 2026)
| Tier | Model | Provider | Input $/MTok | Output $/MTok | Released | Primary use |
|---|---|---|---|---|---|---|
| Frontier-lead (NEW β leader) | claude-opus-5-5 | Anthropic | $4.00 | $20.00 | 2026-09-22 | Supersedes opus-5 and is 20% CHEAPER per token ($4/$20 vs $5/$25; cache read $0.20 = 60% cut; batch $2/$10). #1 AA Intelligence Index v4.3.2 at 58 (max) β a 5-point lead over Fable 5.1 (53) and GPT-6 Astra (53); leads 6 of 10 components incl. AA-Omniscience, GDPval-AA v2.1 (1846), AA-Briefcase (1822), HLE (61.4), SciCode (66.9). 1M ctx / 128K out, cutoff Jun 2026, effort lowβmax (default medium), OpenRouter day-one. ZDR-clean. BUT ~1.6Γ more verbose β AA cost-per-task is FLAT vs opus-5 ($5.98 vs $5.86) β the price cut is not a 20% saving on agentic work. Now watermarked (EU AI Act); thinking cannot be disabled; cyber redirects to opus-4-8. New default for long-horizon agent work, coding, council. β |
| Frontier-lead (superseded β Legacy) | claude-opus-5 | Anthropic | $5.00 | $25.00 | 2026-07-24 | Superseded by opus-5-5 (09-22), which is strictly better AND 20% cheaper. Anthropic moved it to "Legacy (still available)" on the models-overview page while the deprecation table still reads Active / retirement not sooner than 2027-07-24 β no notice, no risk, but no longer a recommended target. RETAINED as the last un-watermarked frontier Claude (matters for anything published as bio-zack) and at any not-yet-swapped pin. Was: #1 AA v4.1 (61 max), joint-#1 Coding Agent Index, GDPval 1861. β |
| Frontier-max (NEW β leader) | claude-fable-5-1 | Anthropic | $10.00 | $50.00 | 2026-09-01 | Supersedes fable-5 at the SAME per-token price, ~25% cheaper EFFECTIVE (75% cache-read cut $1.00β$0.25; batch $5/$25). New #1 AA Intelligence Index (65.7, v4.1.1 β above Opus 5 63.0). 1M ctx / 128K out. First Anthropic model that can DISCOVER vulns (not exploits) β GA cyber path; first Claude with an EU-AI-Act invisible watermark. Still a niche escalation (2Γ opus-5's $5/$25, presumed non-ZDR, watermarked) β opus-5 stays the default + council-synthesis pick. Terminal-Bench-Science 52.6, GDPval 1853. β‘β‘ |
| Frontier-max (superseded) | claude-fable-5 | Anthropic | $10.00 | $50.00 | 2026-06-09 | Superseded by fable-5-1 (09-01, same price, strict upgrade). Never live-routed in vita. SWE-bench Pro 80.3. β‘ |
| Frontier-agent (superseded) | claude-opus-4-8 | Anthropic | $5.00 | $25.00 | 2026-Q2 | Superseded by opus-5 (same price, strictly better). RETAINED as the documented refusal/availability fallback for opus-5 + Fable-5 callers; still live at council + the held opus-4-7-trio's eventual target until swapped. |
| Frontier-reason | gpt-5.5 | OpenAI | $5.00 | $30.00 | 2026-04 | Hardest reasoning, council seat, non-coding strategy |
| Frontier-reason-pro | gpt-5.5-pro | OpenAI | $30.00 | $180.00 | 2026-04 | /council only β rare, highest-stakes (no 5.6-pro exists yet β unchanged) |
| Frontier-6th-gen (watch, NEW) | gpt-6-astra | OpenAI | $10.00 | $50.00 | 2026-09-03 | OpenAI's first 6th-gen model; takes GPT-5.6 Sol's top slot. AA Intelligence Index (v4.1.1) 61 (max) β TIES Sol + Grok-4.6, 5 below Fable 5.1 (66); AA Coding Agent Index 67 (β Opus 5). Real leads in COMPUTER-USE (OSWorld 2.0 72.6, ~2Γ Sol's speed), SWE (Terminal-Bench 4.0 57.7 > Fable 5.1 55.8) and CYBER (first OpenAI model rated CRITICAL; ExploitBench 100% β but exploit-dev gated to Daybreak/BLOCKED in prod API) β all three vendor-harness. 1.05M ctx / 128K out, cutoff 2026-04-30, cached input $1, cache-write $12.50, >272K prompt rebills whole request to ~$20/$75. NOT on OpenRouter (OpenAI-API-only). WATCH, zero routing β $10/$50 = 2Γ Opus 5 / 2Γ Sol for equal-or-lower intelligence; no adopt rationale. Not a council pick (no Omniscience). β |
| Frontier-value (NEW β leader) | gpt-6.1-sol | OpenAI | $2.00 | $10.00 | 2026-09-29 | Supersedes gpt-6-sol at the IDENTICAL $2/$10, seven days later, and it resets the COST frontier rather than the intelligence one. AA v4.3.2: 52 (max) @ $0.72/task Β· 51 (xhigh) @ $0.39 Β· 50 (high) @ $0.32 Β· 48 (medium) @ $0.21 β up from gpt-6-sol's 48. Its high tier is cheaper per task than any tier of any other model in AA's top 25, and it reaches ~90% of opus-5-5's index score (52 vs 58) at ~12% of its per-task cost ($0.72 vs $5.98). 1.1M ctx / 128K out, cutoff Apr 2026, cached input $0.10 (from $1.00); long-context tier >272K rebills at a mild $4/$15 β note that is far gentler than astra's ~$20/$75 whole-request rebill. OpenRouter (openai/gpt-6.1-sol); ChatGPT Work + Codex, not yet main ChatGPT chat. Shipped at DevDay 09-29, ONE DAY before the 09-29 refresh, and was missed. Caveats: all coding/computer-use numbers are OpenAI's own harness (DeepSWE +6.4 vs gpt-6-sol β Astra; AutomationBench +2.2 vs opus-5-5 at ~β
cost; OSWorld 2.0 within 2.1 of Astra at ~1/7 cost/task); openai.com 403'd this run so the vendor announcement is NOT in hand; and the gpt-5.6-luna MRCR 41.3 collapse is still untested across the gpt-6 line β do not trust the nominal 1.1M at depth. Retargets the council OpenAI seat. β¦β¦β¦ |
| Frontier-value (superseded) | gpt-6-sol | OpenAI | $2.00 | $10.00 | 2026-09-22 | Superseded by gpt-6.1-sol (09-29) at the same $2/$10 β AA 48 β 52, cost/task $1.05 β $0.72. Was: same-day answer to Opus 5.5 β OpenAI cut API prices 50%+ 19 days after Astra. $2/$10 (from $4/$20), cached $0.20, 872K ctx (AA), 125 TPS. AA v4.3.2 Intelligence Index 48 (max) β ten points BELOW Opus 5.5 (58) but at ~1/6 the cost-per-task ($1.06 max / $0.13 low vs $5.98). Flag: OpenAI's "beats Opus 5 at 9% of the cost" claim is ONE benchmark (AutomationBench); AA's composite has Sol 48 < Opus 5 51. Supersedes gpt-5.6-sol AND price-dominates gpt-5.5 ($5/$30) β retargets the standing council OpenAI-seat proposal. β¦β¦ |
| Cheap-frontier (NEW) | gpt-6-luna | OpenAI | $0.10 | $0.50 | 2026-09-22 | ~10Γ cheaper than the rostered gpt-5.6-luna ($1/$6); cheaper on output than the budget-fast pin gemini-3.1-flash-lite (~$0.10/$0.40). ~1.05M ctx / 128K out. Do NOT assume the MRCR fix: gpt-5.6-luna's long context COLLAPSED (MRCR 41.3) despite a nominal 1M β re-test before any >100K routing. No independent AA index pulled. WATCH, zero routing until an AA number + an MRCR re-test land. β¦β¦ |
| Frontier-reason (superseded) | gpt-5.6-sol | OpenAI | $5.00 | $30.00 | 2026-07-09 | Superseded by gpt-6-sol (09-22) at $2/$10 β 60%/67% cheaper. Was: direct gpt-5.5 successor at the SAME price, #1 AA Coding Agent Index (80), SOTA Terminal-Bench 2.1 (88.8). The council-seat proposal now targets gpt-6-sol instead. β§ |
| Frontier-reason-value (new) | gpt-5.6-terra | OpenAI | $2.50 | $15.00 | 2026-07-09 | Near-Sol at half price (AA v4.1 55, Coding Agent 77.4 β edges Fable-5's 77.2). Vellum's "default tier". Candidate for future cost-sensitive frontier-reason paths. β§ |
| Cheap-frontier-agent (new) | gpt-5.6-luna | OpenAI | $1.00 | $6.00 | 2026-07-09 | Terminal-Bench 84.7 (> grok-4.5 83.3 > Opus 4.8) at $1/$6. Third arm for the bounded coding pilot. MRCR 41.3% β long-context COLLAPSES despite nominal 1M; never route >100K ctx. β§ |
| Frontier-multimodal | gemini-3.1-pro | $2.00 | $12.00 | 2026-Q1 | Vision-heavy tasks, 1M+ context, benchmarks leader | |
| Frontier-multimodal (watch, NEW β no API id) | gemini-4-argon | $2.00* | $10.00* | 2026-09-30 | Google's first frontier model since February and a real recovery β but UNROUTABLE, and the gate is ACCESS, not evidence. AA v4.3.2 53 (high), #8 of 224, +23 over gemini-3.1-pro-preview (30), back in the top three labs. THERE IS NOTHING TO PIN: no published API model id β Google's post names the model and never gives an API string β and it is absent from OpenRouter, Vertex, Gemini CLI, Cursor and Copilot. Access = Fairwind Program (trusted cyber defenders) only, no GA date; AA benchmarked one provider (Google AI Studio). Genuine firsts: 1M OUTPUT tokens (from 64K, "Long Decode Continuation"), #1 AutomationBench-AA 78% (+7 vs sonnet-5-5 max), 15% AA-Omniscience hallucination β lowest of any model above index 45, and shipped WITHOUT cyber guardrails to defenders (a third posture after Anthropic's refuse-and-fall-back and OpenAI's Daybreak gating). Terminal-Bench 4.0 57 (behind sonnet-5-5 64, opus-5-5 60, Astra 59); AA-Briefcase 1494 Elo / 65% rubric pass (highest recorded); Arena.ai text #1 (1525) but Code Arena web-dev only #8. Vendor: DeepSWE 77.9, LVBench 91.7 (long-video SOTA), CWE-bench v1 68 (tied 1st). *$2/$10 is an UNDATED INTRODUCTORY price β reverts to $4/$20, taking cost/task $1.99 β $3.98 = ~22% ABOVE Astra's $3.26; cached input 95% off (vs Gemini's prior 90%). 62k output tokens/task vs Astra's 27k for equal quality. The 15% hallucination sits on 50% accuracy vs Astra's 63% β it abstains more, it does not know more, so it is NOT an Omniscience-seat challenger. Google's "significantly higher than Astra, Fable and Opus" rests on Vals and does not survive AA (a TIE with Astra, five behind opus-5-5). Cutoff / effort tiers / SynthID / ZDR all UNSTATED β null, not inferred. WATCH, zero routing. ββ | |
| Strong-cheap | kimi-k2.6 | Moonshot | $0.75 | $4.66 | 2026-04-21 | Agent swarms, long-horizon coding, OSS fallback. Retained β its 1M ctx serves the docs-audit / newsletter / consolidator / council-review synthesis roles k2.7-code can't. |
| Strong-cheap-code (pilot) | kimi-k2.7-code | Moonshot | $0.74 | $3.50 | 2026-06-12 | Coding-SPECIALIST successor to k2.6: ~30% fewer thinking tokens, MCP-Mark 81.1. But 256K ctx (NOT 1M) β does not serve k2.6's long-context roles. Vendor-only benchmarks. No current vita site fits (all kimi uses are 1M-synthesis); rostered for future bounded coding agents. ΒΆ |
| Open-frontier (watch) | kimi-k3 | Moonshot | $3.00 | $15.00 | 2026-07-16 | First open-weight(-promised) frontier-tier model (AA v4.1 = 57, #3-4). NOT a k2.6 successor β 4Γ input price puts it in the Terra/grok-4.5/Sonnet band. WATCH: no routing until weights + license land (~Jul-27) and OpenRouter capacity stabilizes (frequent 429s today). β |
| Open-frontier-multimodal (watch, NEW) | qwen3.8-max | Alibaba (Qwen) | $2.00 | $6.00 | 2026-08-03 | First Qwen on the roster. 2.4T sparse-MoE (~95B active), 1M ctx, native text+image+video, #2 Vision Arena / #5 Text Arena. AA Intel Index 53 (search-snippet, unconfirmed β below K3's 57). Vendor-only coding (SWE-bench Pro 67.7). Same $2/$6 band as grok-4.5. WATCH: no routing until an independent SWE-bench number + weights/license (~Aug-10) + capacity + AA vision numbers land. China-jurisdiction risk β non-sensitive only; self-host once weights land. β |
| Strong-cheap-reason | deepseek-v4-pro | DeepSeek | $1.74β | $3.48β | 2026-04-24 | Cheap-frontier reasoning + agentic coding (batch only β 34 TPS). MIT-licensed. |
| Strong-cheap-code | glm-5.2 | Z.ai (Zhipu) | $1.40 | $4.40 | 2026-06-16 | Coding where Opus is overkill, cost-sensitive builds. SWE-bench Pro 62.1 (top open-weights, beats GPT-5.5); usable 1M ctx (IndexShare); high/xhigh thinking. glm-5.1 ($1.05/$3.50) kept as the cheaper fallback for high-volume non-coding passes (feed editors). RETAINED as the pin β glm-5.3 (08-14) supersedes it but has no per-token price/weights yet. Β§ |
| Strong-cheap-code (successor, watch, NEW) | glm-5.3 | Z.ai (Zhipu) | n/a | n/a | 2026-08-14 | Post-training-only refresh of glm-5.2 (same 743B base). Vendor: strongest open-weights coder (+50% vs 5.2; DeepSWE 66.9, AutomationBench 48.2, GDPval-AA v2 1769 β all vendor-harness). Real signal = EMERGENT CYBER (CyberGym 84.5 > Mythos 5 + Sol; 2,436 vulns found across 269 projects) β first roster-adjacent option for AUTHORIZED security-testing where Anthropic refuses, gated on weights + independent repro. WATCH, no routing: per-token price UNPUBLISHED (Coding Plan subs only), weights/license ~2wk out, no OpenRouter, no AA index, coding numbers vendor-only. glm-5.2 stays the pin. β |
| Balanced-lead (NEW β leader) | claude-sonnet-5-5 | Anthropic | $2.00 | $10.00 | 2026-09-28 | Supersedes sonnet-5 at the IDENTICAL $2/$10 (cache read $0.20, cache write $2.50). AA v4.3.2 Intelligence Index 56 (max), #3 of 216 β up from sonnet-5's 38 (#58), the largest same-price gen-over-gen jump in this doc; xhigh 52, high 47. Vendor table BEATS opus-5-5 on Terminal-Bench 4.0 (70.6 vs 66.4) and near-ties it on GDPval 1844/1846, AA-Briefcase 1811/1822, OSWorld 2.1 80.1/81.8, CursorBench 55.5/57.8. 1M ctx / 128K out, cutoff Jun 2026, adaptive thinking, effort lowβmax (default high), 137.6 TPS (fastest Sonnet yet), ZDR-clean, OpenRouter day-one (5 providers), floor 2027-09-28. BUT the cost claims conflict: Anthropic says "up to 30% less per task"; AA measures $7.60/task vs sonnet-5's $5.09 (~49% MORE) and above opus-5-5's $5.98 β "very verbose", 410M output tokens in eval. On AA's frontier opus-5-5 xhigh ties it at 56 for $3.46. Adopt for speed + real-workload token efficiency, NOT for AA cost-per-task. Cyber falls back to sonnet-5; first Sonnet with anti-distillation preserved thinking; watermark status UNKNOWN (not stated at launch). β |
| Balanced (superseded β Legacy) | claude-sonnet-5 | Anthropic | $2.00 | $10.00 | 2026-06-30 | Superseded by sonnet-5-5 (09-28) at the SAME $2/$10 β a strict upgrade for zero per-token cost. Anthropic moved it to "Legacy (still available)" on the models-overview page while the deprecation table still reads Active / floor 2027-06-30 β no notice, no risk, but no longer a recommended target. AA v4.3.2 = 38, #58 of 216, $5.09/task, 74.6 TPS. RETAINED as the documented cyber-fallback target for sonnet-5-5 and at the two pins that cannot take 5.5 (see routing: extractor.py thinking-disable, geo.py DataForSEO namespace). Was: direct sonnet-4-6 successor; $2/$10 intro price made permanent 09-21; new tokenizer emits ~30% more tokens. ΒΆΒΆ |
| Cheap-fast (announced, NOT shipped) | claude-haiku-5-5 | Anthropic | n/a | n/a | β | ANNOUNCED BUT UNSHIPPED β null specs by design. Anthropic's Sonnet 5.5 launch page commits it to "the coming weeks" as the third and final Claude 5.5 model, "built for high-volume and cost-sensitive applications". No model ID, no date, no price, no benchmarks exist at any primary source; the platform model table still lists Haiku 4.5 as the only Haiku. Recorded as a confirmed-forthcoming release with UNKNOWN specs β the 09-21 phantom lesson in its other direction (a primary source exists, so existence is fact; specs are not). Do NOT extrapolate haiku-4-5's $1/$5 β sonnet-5-5 held price exactly while opus-5-5 cut 20%; there is no 5.5-generation pricing rule. Re-run this flow the moment a model card appears. ββ |
| Budget-MoE | minimax-m2.7 | MiniMax | $0.30 | $1.20 | 2026-03-18 | Summarization, synthesis, background research (incumbent β bake-off vs V4-Flash live) |
| Budget-MoE (successor) | minimax-m3 | MiniMax | $0.30 | $1.20 | 2026-05-31 | Direct m2.7 successor at the SAME price: 1M ctx (vs 200K) + native image/video multimodal, MSA sparse attention (~1/20 compute at 1M). Unverified benchmarks. Fold into the budget bake-off. ΒΆΒΆΒΆ |
| Budget-MoE (challenger) | deepseek-v4-flash | DeepSeek | $0.14 | $0.28 | 2026-04-24 | Direct M2.7 challenger: 4x cheaper output, 5x context (1M), Intel Index 47 vs 50. MIT-licensed. PILOT β live on 9 aloop modes. RETAINED as the pin (see v4.1-flash below: not a straight cheaper successor). |
| Budget-MoE (successor, watch, NEW) | deepseek-v4.1-flash | DeepSeek | $0.15β$0.30* | $0.60β$1.20* | 2026-09-10 | 552B MoE (8B active in / 16B out), new Causal Encoder-Decoder architecture, native vision, 1M ctx, up to 384K max output. *PRICING IS TIME-OF-DAY DEPENDENT β a first for this roster. Off-peak $0.15 in / $0.60 out (cache-hit $0.003); PEAK DOUBLES both, peak window MonβFri 01:00β04:00 + 06:00β10:00 UTC. That is 2.1β4.3Γ the incumbent v4-flash's flat $0.28 output, and vita's schedulers run inside those windows β NOT a straight cheaper successor. Pricing is search-snippet, not model-card-confirmed; no AA index. WATCH, zero routing until the peak/off-peak split is modelled against actual job clocks. ββ |
| Budget-fast | gemini-3.1-flash-lite | ~$0.10 | ~$0.40 | 2026-Q1 | High-volume triage: media scoring, daily rollup, trend analysis. RETAINED as the budget-fast pin β gemini-3.7-flash is a pricier mid-tier, NOT a replacement. | |
| Cheap-agentic (watch, NEW) | gemini-3.7-flash | $0.75β$1.50 | $3.75β$7.50 | 2026-08-13 | New mid-tier between flash-lite and gemini-3.1-pro β NOT a flash-lite replacement. Cheap agentic/coding/document workhorse: FrontierCode 1.1 43.6 (vs 3.6-Flash 34.4), GDP.pdf 34, AutomationBench 30.4 (all vendor); 1M ctx / 65K max out, multimodal, OpenRouter day-one. Intro $0.75/$3.75 through 2026-12-31 β $1.50/$7.50 on Jan-1-2027. PRICE DISAGREEMENT: OpenRouter surfaced $0.375/$1.875 β verify at model card. WATCH, no routing until an independent (AA) benchmark + confirmed steady-state price. β | |
| Honesty-specialist | grok-4.20 | xAI | $20.00 | $60.00 | 2026-03-31 | Council seat only β #1 on Omniscience non-hallucination |
| Reasoning-cheap (xAI) | grok-4.3 | xAI | $1.25 | $2.50 | 2026-04-30 | SHADOW PILOT β 5th council seat alongside grok-4.20 until AA publishes Omniscience component. 16-24x cheaper than 4.20, Intel Index 53 (+5), 207 TPS (#1). No model card yet. |
| Frontier-agent-cheap (successor, NEW) | grok-4.7 | SpaceXAI | $2.00 | $6.00 | 2026-09-21 | Supersedes grok-4.6 at the SAME $2/$6 + same 500K ctx. AA-independent: Intelligence Index 46 on v4.3.2 (#16/202; Sol max = 47, Opus 5 max = 51 on the same scale), cost-per-task $3.74, 38.8 TPS, AA: "notably slow and very verbose" (240M output tokens in eval). Vendor table: CursorBench 4.0 46.3 (Sol 41.7, Fable 5.1 51.8), DeepSWE v1.1 71.0 @high (Sol 72.7, Fable 5.1 70.0), Terminal-Bench 4.0 38.0 (Sol 37.3, Fable 5.1 57.9), EEBench 64.0, AA-Briefcase 1657, GDPval 1695 (Fable 5.1 1735). Announcement omits ctx, cutoff, cached price, modality, and whether 4.6's >200K $4/$12 rebill carries β confirm before long-prompt routing. No Omniscience β not council. Refreshes the still-unrun bounded-coding pilot; zero live routing. β |
| Open-frontier-multimodal (watch, NEW β Meta) | muse-spark-1.3 | Meta | $1.25* | $4.25* | 2026-09-02 | First Meta model on the roster. AA v4.3.2 Intelligence Index 48 β above GPT-5.6 Sol (47) and Grok 4.7 (46) β at the LOWEST cost-per-task in the top 12 ($1.60). 1,048,576 ctx, text+image+video in, OpenRouter day-one, open weights "planned". Pricing is search-snippet (mindstudio / openrouter listing), NOT model-card-confirmed. A $0.10/$0.20 "contributor" tier lets Meta TRAIN on prompts/completions (100 RPM cap) β never route vita data there. No independent coding number; WATCH, zero routing* until pricing is confirmed at the card + one coding number lands. Candidate cheap-synthesis / bounded-coding challenger. β |
| Frontier-agent-cheap (superseded) | grok-4.6 | SpaceXAI | $2.00 | $6.00 | 2026-08-12 | Superseded by grok-4.7 (09-21) at the same $2/$6; no formal EOL. Was: supersedes grok-4.5 at the SAME price. +5 AA Intel Index β 61 (claimed GPT-5.6-Sol parity), GDPval-AA v2 1753 Elo (behind only Opus 5, overlapping CIs w/ Fable-5 + Qwen3.8-Max), AA-Briefcase 1577 Elo, cost-per-task $0.84 (ties K3). Lane = agentic knowledge-work + cost-efficiency, NOT raw coding (DeepSWE 65.9 < Sol Max ~73, all xAI-harness). >200K prompt β $4/$12 (whole request rebills); 500K ctx, Feb-1-2026 cutoff; NO Omniscience β not council; vision/tool-use modality unconfirmed. Refreshes grok-4.5's still-unrun bounded-coding pilot; zero live routing swap. AA-index-scale drift flagged (see β). β |
| Frontier-agent-cheap (superseded) | grok-4.5 | SpaceXAI (ex-xAI) | $2.00 | $6.00 | 2026-07-08 | Superseded by grok-4.6 (08-12) at the same $2/$6. Was: cheaper-frontier coding/agentic, AA Intel Index 54 (#4), Coding Agent Index 76, ~2Γ token efficiency. Beats Opus 4.8 on Terminal Bench 2.1 (83.3) + SWE Marathon (29.0); trails on SWE-bench Pro (64.7). 500K ctx; NO Omniscience. If ever adopted from the pilot slot, prefer grok-4.6. β¦ |
β V4-Pro on a 75% promo through 2026-05-05: $0.435/$0.87 per MTok. Standard pricing resumes May 6. Cache-hit input is $0.145.
β‘claude-fable-5 (GA 2026-06-09): adaptive thinking always-on (no budget_tokens, no thinking:{type:disabled} β both 400), raw chain-of-thought never returned (summaries via thinking.display), safety classifiers can return stop_reason:refusal (opt-in fallback to opus-4-8 required in any caller), and 30-day data retention required β NOT available under ZDR. Reserve for highest-stakes single calls; 2Γ Opus output cost. Sibling claude-mythos-5 (identical specs, no classifiers) is Project-Glasswing-only β not adoptable by vita. Bench note (UPDATED 2026-07-10): SWE-bench Pro 80.3 confirmed (official + Vellum); AA Intelligence Index rescaled to v4.1 β Fable-5 (max) = 60, still #1, one point above gpt-5.6-sol (direct-confirmed at artificialanalysis.ai this refresh; supersedes the "~65" carried from the old index version β the two scales are NOT comparable). SWE-bench Verified ~95 remains single-source-aggregation, treat as indicative. New comparative datapoints from the 5.6 launch: Fable-5 keeps decisive leads on SWE-bench Pro (80 vs Sol 64.6), AA-Briefcase (rubric 56% vs 42%, analytical Elo 1764 vs 1592), GDPval-AA-v2 and HealthBench Professional; Sol beats it on Coding Agent Index (80 vs 77.2), Terminal-Bench 2.1 (88.8 vs 86.0) and Agents' Last Exam (53.6 vs 40.5) at ~1/3 the cost-per-task. Availability event (verified 2026-07-05): Fable 5 + Mythos 5 were export-control-SUSPENDED by a US Commerce Dept directive on 2026-06-12 (barring access by any foreign national, globally) β a full shutdown β and RESTORED 2026-07-01 after the directive was lifted. The slot now carries a standing geopolitical-availability risk on top of the refusal caveat: any Fable-5 caller needs the opus-4-8 fallback wired for BOTH refusal AND availability. (The 06-23 review saw "Fable-5 pulled by export control" and dismissed it as an unverified aggregator rumor β it was real. Do not auto-dismiss model-status rumors; verify.)
β‘β‘claude-fable-5-1 (Anthropic, GA 2026-09-01, API id claude-fable-5-1): direct successor to fable-5 at IDENTICAL $10/$50 input/output, but ~25% cheaper EFFECTIVE (up to ~45% on complex/agentic coding) from token-efficiency gains plus a 75% cache-read cut ($1.00 β $0.25/MTok); batch $5/$25. 1M ctx / 128K max output, adaptive thinking always-on (same integration constraints as fable-5: no budget_tokens, no thinking-disable). AA-independent: Intelligence Index (2026 / v4.1.1 scale) = 65.7 β new single #1, above Opus 5 63.0 and Fable 5 62.1 (do NOT compare to the old v4.1 numbers β different scale). Anthropic's own launch table: Terminal-Bench-Science 0.1 52.6% (Opus 5 29.0, Fable 5 24.7, GPT-5.6 Sol 22.4); Terminal-Bench 4.0 agentic-coding 55.8% (Fable 5 42.0; Mythos 5.1 60.9); CursorBench 3.2.0 73.4; HLE 65.0 w/tools; OSWorld 2.0 strict 41.7; AutomationBench 31.4; GDPval-AA v2 1853 (Opus 5 1824, Fable 5 1723). Cyber (genuinely new): strongest cyber capability of any Anthropic model to date; Claude Code sees ~60% fewer cyber-safeguard interventions per session; and fable-5-1 can now discover software vulnerabilities but NOT develop exploits β the first GA, generally-reachable Anthropic option for authorized/defensive security-testing (opus-5/opus-4-8/fable-5 all refuse; glm-5.3's cyber was vendor-only + weights-not-landed; Mythos 5.1 is trusted-access). Watermark (new integration consideration): first Claude models to embed an invisible numerical text/file watermark (EU AI Act, models released after 2026-08-02), detectable only via Anthropic's detection API β flag for anything vita publishes as bio-zack. Biology safeguards fire 85% less for benign requests; alignment note β significantly less likely than Mythos 5 to access resources outside its test environment. ZDR/retention NOT re-confirmed this run (fable-5 required 30-day retention / no ZDR β verify before routing sensitive paths). GA as claude-fable-5-1 on Anthropic API + AWS/Bedrock + Google Cloud + Microsoft Azure/Foundry + desktop app. Sibling claude-mythos-5-1 = same model, cyber/life-sciences safeguards, trusted-access US-orgs-only β NOT vita-adoptable (same as Mythos 5). Sources fetched this run: anthropic.com/claude-fable-and-mythos-5-1, marktechpost.com (09-01), benchlm.ai/benchmarks/artificialanalysis. Routing: no live pin moves β fable-5-1 remains a per-task niche escalation for the hardest builds, not a default (opus-5 is cheaper + ZDR-clean + un-watermarked).
βclaude-opus-5-5 (Anthropic, GA 2026-09-22, API id claude-opus-5-5, alias identical, Bedrock/Vertex/Foundry anthropic.claude-opus-5-5): direct successor to opus-5 and the first model in this doc's history that is simultaneously the independent #1 AND cheaper per token than the incumbent live default. Pricing $4/$20 (opus-5 was $5/$25), 5m cache write $5, 1h cache write $8, cache read $0.20 β note this is a 0.05Γ multiplier, a THIRD distinct cache shape after the standard 0.1Γ and Fable 5.1/Mythos 5.1's 0.025Γ; batch $2/$10; fast mode $8/$40 (vs opus-5's $10/$50). 1M ctx / 128K max output (300K on the Batch API with the output-300k-2026-03-24 beta header), reliable knowledge cutoff Jun 2026, effort low/medium/high/xhigh/max with medium the DEFAULT (opus-5's callers that never set effort will land on medium β set it explicitly if you were relying on opus-5's behavior), retirement floor 2027-09-22, tool-use system prompt overhead a flat 286 tokens for auto/none (opus-5 charged 406 for any/tool). Live on Claude API + Claude Platform on AWS + Bedrock + Google Cloud + Azure/Foundry + OpenRouter day-one (anthropic/claude-opus-5.5). AA-independent (model page + launch article both direct-fetched): Intelligence Index v4.3.2 = 58 (max), #1 of 212 β 54 at high, 51 at medium; four of five effort tiers sit on AA's Pareto frontier. It leads six of the ten index components: Humanity's Last Exam 61.4% (previous best 59.1), SciCode 66.9% (prev 63.1), GDPval-AA v2.1 1846 Elo (+111 vs Fable 5.1, +138 vs Opus 5), AA-Briefcase v1.1 1822 Elo (+143 vs Fable 5.1), AA-Omniscience, and AutomationBench-AA. Behind on CritPt, AA-LCR, GDP.pdf. Anthropic's own table: Terminal-Bench 4.0 66.4 (Fable 5.1 55.8 / Opus 5 52.3), CursorBench 4.0 57.8 (51.8 / 46.6), FrontierCode v1.1 54.4 (50.3 / 48.0), GDPval-AA v2.1 1846 (1735 / 1708), OSWorld 2.0 81.8 (80.7 / 74.0), Chartography 89.0 (88.4 / 83.4), Terminal-Bench-Science 0.1 58.7, HLE-with-tools 67.7. Output speed +30% vs opus-5. THE COST TRAP β record this before any cost estimate: AA measures cost-per-task at $5.98, essentially FLAT against opus-5's $5.86, because 5.5 emits ~1.6Γ the output tokens (~119k per index task; 260M across AA's eval vs an 88M median β AA's label is "very verbose"). AA states it plainly: "maintains cost parity with Opus 5 despite 1.6x higher token usage." So the 20% per-token cut buys a large intelligence gain at roughly the SAME agentic spend. The genuine savings are on cache-heavy and short-output paths, where the 60% cache-read cut dominates and the verbosity penalty doesn't apply. BENCHMARK DISAGREEMENT (both recorded, flagged): Terminal-Bench 4.0 β Anthropic reports 66.4% and puts GPT-6 Astra at 57.9%; AA's independent run reports 59.6% and calls it a TIE with Astra. Different harnesses. AA also notes its AutomationBench runs were performed without fallback models, so safeguard interventions counted as outright failures β an artificial penalty. Prefer AA for cross-vendor comparison, Anthropic for gen-over-gen. HLE likewise differs by condition (AA 61.4 index-component vs vendor 67.7 with tools) β not a contradiction. Safety / integration: ZDR available (keeps opus-5's advantage over Fable 5.1's mandatory retention); best automated-behavioral-audit score of any Anthropic model to date; 85% fewer containment-boundary circumvention attempts; tied Fable 5.1 for the lowest prompt-injection success rate Gray Swan has recorded (independent). Cyber: safeguards similar to Fable 5.1, and it REDIRECTS most cyber tasks to opus-4-8 β keep that fallback wired; NOT an authorized-security-testing option (Fable 5.1 keeps the discover-not-exploit GA path). Thinking can no longer be disabled (same constraint as Fable 5.1 β no thinking:{type:disabled}; note this breaks the exact workaround the 09-21 haikuβsonnet-5 swap needed at feed/constitutions/extractor.py, so do NOT port that pattern to opus-5-5 on a small-max_tokens path). NEW TRAP: a "preserved thinking anti-distillation safeguard" is active for API accounts created after 2026-08-31 β verify which vita account that hits before assuming opus-5 thinking-block behavior carries. WATERMARK: it embeds the EU-AI-Act invisible watermark, same as Fable 5.1. opus-5 was the last un-watermarked frontier Claude β so the watermark consideration for anything vita publishes AS bio-zack (X, email, PR text, reports) now applies to the Opus line too, and opus-5 is the model to reach for if that matters. Sources fetched this run: platform.claude.com models/overview + pricing + model-deprecations, anthropic.com/news/claude-opus-5-5, artificialanalysis.ai/models/claude-opus-5-5, artificialanalysis.ai/articles/claude-opus-5-5, vellum.ai/blog/claude-opus-5-5-benchmarks-explained, openrouter.ai/anthropic/claude-opus-5.5, techcrunch.com (09-22), officechai.com.
β¦β¦gpt-6-sol + gpt-6-luna (OpenAI, GA 2026-09-22 β the same day as Claude Opus 5.5, 19 days after Astra): OpenAI cut API prices 50%+ across the new tier. gpt-6-sol $2/$10 (down from the Astra-era $4/$20), cached reads $0.20 (90% discount), 872K context per AA, 125 TPS at max. AA v4.3.2 Intelligence Index = 48 (max), cost-per-task $0.13 (low) β $1.06 (max). Read that against Opus 5.5's 58 at $5.98/task: Sol is ten index points lower at roughly one-sixth the cost per task β a value play, not an intelligence leader. Vendor-vs-independent gap flagged: launch coverage repeats OpenAI's claim that Sol "beats Opus 5 at 9% of the cost per task" β that is ONE benchmark (AutomationBench, Sol xhigh 33.2% vs Opus 5 max 26.9%, Fable 5.1 31.4%); AA's ten-eval composite has Sol 48 < Opus 5 51 < Opus 5.5 58. Do not carry the vendor framing. It supersedes gpt-5.6-sol and price-dominates gpt-5.5 ($5/$30, still the live council OpenAI seat) β which makes the standing 07-10 "gpt-5.5 β gpt-5.6-sol" council proposal obsolete; retarget to gpt-6-sol. gpt-6-luna $0.10/$0.50 (down from $0.20/$1.20), ~1,050,000 ctx / 128K out β about a 10Γ cut against the rostered gpt-5.6-luna ($1/$6) and cheaper on output than the budget-fast pin gemini-3.1-flash-lite (~$0.10/$0.40), which makes it the first credible challenger in that band in months. Hard caveat: do NOT assume the long-context fix. gpt-5.6-luna posted MRCR 41.3 β a long-context collapse despite a nominal 1M window β and no MRCR number was pulled for gpt-6-luna. Re-test before any >100K-context routing. Neither model was fetched at an OpenAI model card this run (Sol's AA release page + search coverage only); no independent AA index for Luna. Both rostered; Sol is a live proposal target (council seat), Luna is WATCH/zero-routing until an AA index + an MRCR re-test land. Sources fetched this run: artificialanalysis.ai/models/releases/gpt-6-sol, venturebeat.com (09-22), vellum.ai, llm-stats.com/models/gpt-6-luna, openrouter.ai/openai/gpt-6-luna, digitalapplied.com.
ββdeepseek-v4.1-flash (DeepSeek, GA 2026-09-10): 552B-parameter MoE activating 8B for input / 16B for output on a new Causal Encoder-Decoder architecture, native image input, 1,048,576 ctx, up to 384K max output. Coverage credits it with cutting agent-memory costs ~4Γ. The reason it is NOT a drop-in successor to the rostered deepseek-v4-flash: pricing varies by time of day, which nothing else on this roster does. Off-peak: $0.15/MTok cache-miss input, $0.003/MTok cache-hit input, $0.60/MTok output. Peak rates are DOUBLE, and the peak window is MonβFri 01:00β04:00 and 06:00β10:00 UTC. The incumbent v4-flash is a flat $0.14/$0.28 β so v4.1-flash's output runs 2.1Γ off-peak to 4.3Γ at peak, and vita's scheduled jobs (nightly rollup 00:05 PT = 07:05 UTC, /sleep 00:30 PT = 07:30 UTC) land squarely inside the second peak window. A naive swap would cost MORE on exactly the jobs that drive v4-flash's volume. All pricing is search-snippet (openrouter/llm-stats/dataconomy/techbriefly), not model-card-confirmed; no AA Intelligence Index pulled. WATCH, zero routing until (a) the model card confirms the peak/off-peak schedule and (b) someone models it against .aloop job clocks. Also note this partly answers the long-standing "DeepSeek V4 watch" β the V4 line is shipping point releases, not a V4 successor generation. Sources fetched this run: openrouter.ai/deepseek/deepseek-v4.1-flash, llm-stats.com/models/deepseek-v4.1-flash, dataconomy.com (09-11), techbriefly.com (09-11).
β
gpt-6-astra (OpenAI, GA 2026-09-03, API id gpt-6-astra): OpenAI's first 6th-gen model, replacing GPT-5.6 Sol at the top of the lineup. Pricing $10/$50 input/output, cached input $1, cache-write $12.50; fast mode = 2Γ speed at 2Γ price; above 272,000 input tokens the ENTIRE request reprices to ~$20 input / ~$75 output (input+cache 2Γ, output +50%). Context 1,050,000 (max input ~922K, max output 128K), knowledge cutoff 2026-04-30, reasoning effort low/medium/high/xhigh/max (no none). NOT on OpenRouter this run β OpenAI-API-only at launch; rolled out FIRST to the Daybreak application-only cybersecurity cohort, then Plus/Pro/Business/Enterprise, then API. AA-independent (article direct-fetched): Intelligence Index (v4.1.1/"2026" scale) 61 β "equal to GPT-5.6 Sol in the Index at 61," 5 points below Fable 5.1; AA Coding Agent Index 67 ("approximately equal to Claude Opus 5"); "70% more token-efficient" than Sol on coding but "75% more expensive per task at max effort"; hallucination rate decreased 92%β51% at max effort (still high); AA-Briefcase +~80 Elo but GDPval-AA v2 β~80 Elo vs Sol (a DROP). AA-index tier caveat: the "61 = Sol" tie is AA's MAX-effort read; this doc's v4.1.1 baseline (line ~98) has Sol 58.9 / Fable 5.1 65.7 on a default/blended read β effort-tier ambiguity, both recorded, do NOT overwrite the baseline. One aggregator listed 61.2 vs AA's 61 (minor, both recorded). Vendor/OpenAI-harness benchmarks (no independent reproduction): FrontierMath Tier 4 97.6%, DeepSWE v1.1 74.1%, ARC-AGI-3 99.9%, OSWorld 2.0 72.6% at ~40min/task (vs Sol 65.7% @ ~75min β computer-use SOTA), Terminal-Bench 4.0 57.7% (vs Sol 37.3, Fable 5.1 55.8 β OpenAI calls it best-SWE-to-date), Terminal-Bench Science 64.6%, SRE-Bench pass@1 88%, Agents' Last Exam 59.3%, HLE-with-tools 57.2% (trails Fable 5.1), FrontierCode Extended 64.5% (trails Fable 5 64.9). Cyber (the vita-relevant TRAP): FIRST OpenAI model rated CRITICAL for cybersecurity under the Preparedness Framework (also HIGH bio/chem); ExploitBench 100% without production safeguards (Sol 78.5%) β BUT exploit-CREATION ships GATED behind the Daybreak program and is BLOCKED in the production API (TheHackerNews: "OpenAI Blocks PoC Exploit Requests"), so the reachable production behavior for authorized security-testing is still a refuse, exactly like the rest of the roster β Astra does NOT displace Fable 5.1's new discover-not-exploit GA path as vita's authorized-security option. Not researched this run (pull before any routing): ZDR/data-retention posture, watermark (EU-AI-Act) status, vision/OCR modality, any independent (non-AA) SWE-bench Pro/Verified reproduction. GPT-5.6 Sol NOT deprecated β no EOL; Sol is on a promo ($4 input / $0.40 cached / $20 output) through 2026-11-21. Routing: WATCH, zero live pin moves β $10/$50 is 2Γ Opus 5's $5/$25 and 2Γ Sol's $5/$30 for equal-or-lower measured intelligence; the differentiated lanes are computer-use + gated cyber, neither of which vita routes to today. Sources fetched this run: artificialanalysis.ai/articles/benchmarking-gpt-6-astra, llm-stats.com/models/gpt-6-astra, digitalapplied.com/blog/gpt-6-astra-price-benchmarks-guide, thehackernews.com (09-2026 ExploitBench), unite.ai, marktechpost.com (09-03), officechai.com, trendingtopics.eu; official openai.com/index/gpt-6-astra 403'd on direct fetch (facts corroborated via the above + the search summary of the official page).
Practitioner signal (bio-zack, 2026-09-22, not a benchmark): he uses astra as his go-to ADVERSARIAL CODE REVIEWER and is rate-limited on it β "went through 2 banked resets with astraβ¦ need this goddamn reset so i can do my adversarial review passes on the last 20k LOC. astra so good for that" (data/signal/cullen.jsonl, 2026-09-22T15:50Z). "banked resets" implies a subscription plan, not the API, so this does NOT touch the $10/$50 API cost math above. What it bears on is quality: the next review should weigh astra for vita's adversarial-verify/code-review seats (oneshot reviewers, council) before holding WATCH on price alone. Unadjudicated β his usage, not a routing request.
βclaude-opus-5 (Anthropic, GA 2026-07-24, API id claude-opus-5): direct successor to opus-4-8 at IDENTICAL $5/$25 (Anthropic: "greatly improved performance for the same cost as Opus 4.8"; +22% on hardest agentic coding vs prior Opus, organic-chem +10.2pts, protein +7.7pts, Frontier-Bench v0.1 more than doubles 4.8). AA (direct-fetched artificialanalysis.ai + AA article + AA on X): Intelligence Index v4.1 = 61 (max) β narrowly the single most intelligent model, edging Fable-5 60 / Sol 59 / K3 57; ~26% lower Cost-per-Task than Fable-5 ($17.79 vs $22.30 at max; at HIGH effort $10.41/task BEATS Fable-5 by 32 Elo at 47% of its cost); joint-first AA Coding Agent Index (run in Claude Code @ xhigh) + highest SWE-Atlas-QnA; GDPval-AA v2 1861 Elo (>100 ahead of Fable-5/Sol); AA-Briefcase 1720 Elo (+146 vs Fable-5). Anthropic's own: CursorBench 3.2 within 0.5% of Fable-5 at half cost; ARC-AGI-3 ~3Γ next-best; OSWorld-2.0 (computer use) surpasses Fable-5 at ~1/3 cost; Zapier AutomationBench 1.5Γ next-best at same cost/task. New default on Claude Max, strongest on Claude Pro; 2Γ fast-mode (~$10/$50, ~2.5Γ speed). Context 1M per launch coverage (BigGo/AI-Weekly/MarkTechPost) β official anthropic.com/news + system-card blurb stated PRICE only this run; treat 1M as secondary-source until confirmed at the model card. Safety: NO data-retention requirement (ZDR-clean β a real advantage over Fable-5's mandatory 30-day/no-ZDR). Cyber classifiers intervene ~85% less often than Fable-5 but the model still blocks binary vuln-scanning / pentest / exploit-gen and can return refusal stops β Anthropic ships an opus-4-8 fallback for flagged requests; wire it in any opus-5 caller (same discipline as Fable-5, minus the availability risk). Remains BEHIND Mythos-5 (Project-Glasswing-only, not vita-adoptable) on cyber-exploitation β do not route offensive-security here. Posts the LOWEST score on Anthropic's automated behavioral-audit scale of any recent Claude (best-behaved). NO independent SWE-bench Pro/Verified head-to-head vs Fable-5 pulled this run (AA leads are Intelligence Index / GDPval / AA-Briefcase / Coding-Agent-Index); OpenRouter availability unverified this run (Anthropic API day-one). Sources: anthropic.com/news/claude-opus-5, the Opus-5 system card PDF, artificialanalysis.ai/articles/claude-opus-5-leader-agentic-knowledge-work, x.com/ArtificialAnlys.
Β§glm-5.2 (Z.ai/Zhipu, GA 2026-06-16, 753B open weights, MIT): SWE-bench Pro 62.1 > GPT-5.5 58.6 > glm-5.1 58.4; FrontierSWE 74.4% and MCP-Atlas 77.0 near-tie Opus 4.8 (75.1 / 77.8). 1M usable ctx via IndexShare (~2.9Γ FLOP cut at full length), high/xhigh thinking-effort levels, TTFT ~10.8s (async-friendly, NOT chat-latency). Pricing / context / release-date direct-verified this refresh (openrouter.ai + llm-stats.com agree); benchmarks are from the Z.ai launch coverage (VentureBeat) and are NOT yet on artificialanalysis.ai β re-confirm next refresh, and AA Intelligence Index is unpublished (general-reasoning score held = glm-5.1). China data-jurisdiction risk via API/OpenRouter, but MIT weights allow self-host; vita already scopes glm to non-sensitive code paths, so OpenRouter routing stays in policy. OpenRouter lists it text-in/out β not a vision/OCR model.
ΒΆkimi-k2.7-code (Moonshot, GA 2026-06-12, 1T total / 32B active MoE, Modified-MIT open weights, OpenRouter/DeepInfra $0.74/$3.50 β official Kimi API $0.95/$4.00; one aggregator lists $1.14/$4.80 outlier): coding-SPECIALIZED variant (NOT a general K2.7), forced thinking mode, ~30% fewer reasoning tokens than K2.6. Context 262K (256K) β a hard drop from k2.6's 1M. ALL benchmarks are Moonshot's own proprietary suites (Kimi Code Bench v2 +21.8% / Program Bench +11.0% / MLS Bench Lite +31.5% / MCP-Mark Verified 81.1 β all vs K2.6); devops.com confirms no independent third-party results on SWE-bench Verified/Pro, LiveCodeBench, or GPQA as of release. The SWE-bench Pro 58.6 / Verified 60.4 floating on aggregator pages are vendor-reported and suspect (58.6 == K2.6's exact number). Same "vendor-only, re-confirm at AA" posture as glm-5.2. text+image input.
βclaude-sonnet-5-5 (Anthropic, GA 2026-09-28, API id claude-sonnet-5-5, alias identical, Bedrock/Vertex/Foundry anthropic.claude-sonnet-5-5): direct successor to sonnet-5 and the second model in the Claude 5.5 family. Pricing is IDENTICAL to sonnet-5 β $2/$10, cache read $0.20, cache write $2.50 (batch 50% off per the platform pricing footnote, not separately confirmed at launch). 1M ctx / 128K max output (300K on the Batch API with the output-300k-2026-03-24 beta header), reliable knowledge cutoff Jun 2026, adaptive thinking with effort lowβmax and high the DEFAULT (note: a higher default than opus-5-5's medium β a caller that sets no effort gets more thinking here than on the flagship), retirement floor 2027-09-28. Live on Claude API + Claude Platform on AWS + Bedrock + Google Cloud + Azure/Foundry + OpenRouter day-one (anthropic/claude-sonnet-5.5, 5 providers). Text + image input per the Anthropic model table (an OpenRouter snippet additionally claims file/PDF input β not model-card-confirmed, verify before routing PDFs). AA-independent (model page + leaderboard both direct-fetched): Intelligence Index v4.3.2 = 56 (max), #3 of 216 β 52 at xhigh, 47 at high; sonnet-5 on the same scale is 38, #58, so this is an 18-point jump at zero per-token cost, the largest same-price generational move this doc has recorded. Output speed 137.6 TPS ("well above average", fastest Sonnet to date). Anthropic's own table (Sonnet 5.5 / Sonnet 5 / Opus 5.5 / GPT-6 Sol): Terminal-Bench 4.0 70.6 / 10.3 / 66.4 / β; CursorBench 4.0 55.5 / 34.1 / 57.8 / β; FrontierCode 1.1 46.2 / 42.4 / 54.4 / 49.3; GDPval-AA v2.1 1844 / 1449 / 1846 / 1487; AA-Briefcase v1.1 1811 / 1359 / 1822 / 1483; HLE-with-tools 64.5 / 54.9 / 67.7 / β; OSWorld 2.1 80.1 / 57.0 / 81.8 / β; Chartography 61.6 / 15.6 / 64.4 / 53.6. Read Terminal-Bench 70.6-vs-66.4 carefully β it is the cheaper model beating the flagship on Anthropic's own agentic-coding harness, which is a real signal, but the 09-23 precedent is that AA's independent Terminal-Bench run read ~7 points below Anthropic's for opus-5-5. The sonnet-5 baseline of 10.3 is implausibly low as a pure capability delta (a 7Γ gen-over-gen jump on one benchmark, against AA's +18 composite) β read it as "sonnet-5 largely fails to drive the TB4 harness", not as a capability ratio. THE COST TRAP β the two sources point in OPPOSITE directions and both are recorded: Anthropic claims "costs up to 30% LESS per task" than sonnet-5, backed by customer workloads (Balyasny ~121K tokens/answer vs 497K on Sonnet 5; Base44 3.6 build iterations vs 7.7 on Opus 5; Zendesk 20% faster ticket processing). AA measures the opposite: $7.60 per task vs sonnet-5's $5.09 β ~49% MORE, and above opus-5-5's $5.98, with the label "notably fast, however very verbose" (410M output tokens across the eval vs sonnet-5's 370M). These are not reconcilable from the outside and neither should stand alone: AA runs a fixed hard-reasoning set at max effort, where a more capable model thinks longer for the same task; Anthropic measures iteration and tool-call counts on real work, where fewer round-trips dominate. And on AA's own effort frontier the flagship dominates it: opus-5-5 at xhigh scores the identical 56 for $3.46 (2.2Γ cheaper per task), while opus-5-5 medium (51 @ $1.34) beats sonnet-5-5 high (47 @ $1.08) on intelligence for 26Β’ more. So the adopt rationale is NOT cost β it is a very large capability jump for zero per-token change, plus genuine speed. Measure tokens on any path you move. Safety / integration: ZDR β the launch page states zero data retention, so it keeps the posture opus-5/opus-5-5 have and Fable 5.1 lacks. First Sonnet to launch with cyber safeguards and fallbacks: higher-risk cyber tasks VISIBLY FALL BACK TO claude-sonnet-5 β a different fallback target than opus-5-5's opus-4-8 redirect, so keep sonnet-5 reachable on any path that might brush cyber; tiered access via the Cyber Verification Program. Biology safeguards unchanged from sonnet-5. First Sonnet deployed with anti-distillation classifiers + expanded preserved thinking ("Claude's thinking cannot be decoupled from the account that created it") β same class of trap flagged for opus-5-5 on 09-23; verify which vita API account that hits before assuming sonnet-5 thinking-block behavior carries. THINKING-PARAMETER TRAP (this is the one that bites vita today): sonnet-5-5 is adaptive, and adaptive-thinking Claudes reject thinking:{type:disabled} with a 400 (fable-5 and fable-5-1 both do; opus-5-5's thinking "can no longer be disabled"). scripts/feed/constitutions/extractor.py:72 passes exactly that param β the workaround the 09-21 haikuβsonnet-5 swap needed β so that site cannot take sonnet-5-5 as a one-line swap; see the routing table. WATERMARK: NOT MENTIONED on the launch page β status UNKNOWN, recorded as unknown rather than inferred. The stated EU-AI-Act rule covers "models released after 2026-08-02", which would imply it carries, but that is inference and not a fetched fact β confirm before routing anything published AS bio-zack here. Sources fetched this run: anthropic.com/claude-sonnet-5-5, platform.claude.com models/overview + model-deprecations, artificialanalysis.ai/models/claude-sonnet-5-5, artificialanalysis.ai/models/claude-sonnet-5, artificialanalysis.ai/leaderboards/models, marktechpost.com (09-28). Note anthropic.com/news/claude-sonnet-5-5 404s β the canonical URL drops the /news/ segment.
ββclaude-haiku-5-5 (Anthropic, ANNOUNCED 2026-09-28, NOT SHIPPED): the only primary fact is one sentence on Anthropic's Sonnet 5.5 launch page β "Claude Haiku 5.5, built for high-volume and cost-sensitive applications, will join the Claude 5.5 family in the coming weeks." No model ID, no release date, no pricing, no context window, no benchmarks exist at any primary source. Verified absent from the platform.claude.com model-comparison table (which still lists claude-haiku-4-5-20251001 as the current and only Haiku) and from the deprecation table. Secondary coverage (kucoin, zeniteq, orcarouter, cellcog) adds no primary facts; one item references an earlier "Haiku 5.5 is dead" leak that this announcement falsifies. Handled per the standing rumor rule, in its less-common direction: the 09-21 lesson is quarantine unverified rumors, never assert absence from them β here a primary source exists, so its existence as a committed release is a FACT, while every spec is genuinely UNKNOWN. Rostered with null specs; not scored in the domain grid (scoring, or carrying haiku-4-5's numbers forward, would be fabrication). Do NOT assume it inherits haiku-4-5's $1/$5 or 200K context: sonnet-5-5 held sonnet-5's price exactly while opus-5-5 CUT opus-5's by 20%, so there is no consistent 5.5-generation pricing rule to extrapolate from. Practical read for vita: this is the answer to the haiku-4-5 succession question, but it has not arrived β and in the meantime the question has gone quiet on its own, because vita has zero live haiku-4-5 pins (the three were routed to sonnet-5 / gemini-3.1-flash-lite on 09-21). Re-run this flow the moment a model card appears.
ΒΆΒΆclaude-sonnet-5 (Anthropic, GA 2026-06-30, SUPERSEDED by sonnet-5-5 on 2026-09-28 at the identical $2/$10; moved to "Legacy (still available)" on the models-overview page while the deprecation table still reads Active / floor 2027-06-30 β both recorded, no notice, no risk): direct successor to sonnet-4-6 and Anthropic's new default on Free/Pro. Pricing CORRECTED 2026-09-21 (direct-fetched platform.claude.com/docs/en/about-claude/pricing): $2/$10 is now the STANDARD price β "The $2/$10 ... announced at launch as introductory pricing through August 31, 2026, is now the standard price. The previously scheduled increase to $3/$15 ... on September 1, 2026 will not occur." Cache write $2.50 (5m) / $4 (1h), cache read $0.20, batch $1/$5, 1M ctx / 128K out, reliable cutoff Jan 2026, retirement not sooner than 2027-06-30. CRITICAL for cost math: an updated tokenizer (Claude 4.7+) produces "approximately 30% more tokens for the same text" (official) β so the 33% per-token discount vs sonnet-4-6 ($3/$15) roughly cancels; treat as cost-NEUTRAL vs 4.6, not cheaper. (Superseded text: the earlier ~1.0-1.35Γ estimate and "rises Sep 1".) Perf "close to Opus 4.8", "substantial improvement over Sonnet 4.6" β Anthropic gives qualitative claims, no headline numbers (pull AA/SWE-bench next refresh). Safety note: improved agentic safety vs 4.6 but "substantially poorer on cybersecurity tasks vs Opus" β do not route security/vuln work here. Context window not stated in the announcement β confirm before long-context routing. Pilot-with-voice-eval, not an auto-swap: voice-critical paths (coaching, podcast transcripts, silicon-zack chat) need a drift check first.
ΒΆΒΆΒΆminimax-m3 (MiniMax, GA 2026-05-31, first open-weights model combining frontier coding + 1M ctx + native multimodality): OpenRouter flat $0.30/$1.20 (same as m2.7) with prompt-caching cutting effective cost 60-80% on repeated context. Native text+image+video input (vs m2.7 text-only), 1M context (vs m2.7's 200K), MiniMax Sparse Attention (MSA β KV-block selection, ~1/20 per-token compute at 1M). Pricing disagreement recorded: OpenRouter flat $0.30/$1.20 vs direct-API $0.60/$2.40-standard + launch-promo + a 512K long-context band that DOUBLES the rate β verify whether OpenRouter's rate is promo or permanent before high-volume routing. Benchmarks unverified (no AA numbers pulled this run; launch coverage claims GPT-5.5/Opus-4.7-tier coding β vendor framing). Missed by the last several refreshes (released May 31; doc last refreshed Jun 23).
β§gpt-5.6 family (OpenAI, GA 2026-07-09; Sol previewed since 06-26; SiliconANGLE puts broad rollout completing 07-10 after a US Commerce Dept pre-release review β the Fable-5 export-suspension pattern is now apparently a standing regime for frontier launches). Model IDs gpt-5.6-sol / -terra / -luna; bare gpt-5.6 routes to Sol. ALL tiers: 1M ctx, 128K max output, knowledge cutoff 2026-02-16, six reasoning-effort levels (none/low/medium/high/xhigh/max), all live on OpenRouter day-one (Terra listing direct-verified: $2.50/$15, 1M ctx, text+image). NEW Responses-API features: programmatic tool calling (model-written JavaScript orchestrates tool calls in an isolated no-network V8 β 38β63.5% token reductions claimed), native subagent spawning, explicit prompt-cache breakpoints (30-min minimum cache life; cache writes now billed 1.25Γ uncached input, reads keep the 90% discount: Sol $0.50 / Terra $0.25 / Luna $0.10 cached reads). OpenAI claims 54% token-efficiency gain vs prior gen. Benchmarks (AA direct-fetched + Vellum): AA Intelligence Index v4.1 Sol 59 / Terra 55 / Luna 51 (Fable-5 = 60, #1); Coding Agent Index Sol 80 #1 / Terra 77.4 / Luna 74.6 (Fable-5 77.2); Terminal-Bench 2.1 Sol 88.8 SOTA (91.9 "Ultra") / Terra 87.4 / Luna 84.7 (Fable-5 86.0); Agents' Last Exam Sol 53.6 / Terra 50.4 / Luna 50.3 (Fable-5 40.5); MRCR Sol 91.5 / Terra 89.6 / Luna 41.3 β long-context disqualified; AA cost-per-task Sol $1.04 (~1/3 Fable-5) / Terra $0.55 / Luna $0.21. SWE-bench Pro: Sol 64.6 vs Fable-5's 80 β OpenAI disputes the benchmark's validity and omitted SWE-bench Verified/GPQA/AIME/MMLU/FrontierMath where Claude leads (selective reporting β weigh vendor-favorable numbers accordingly). METR caveat: multiple launch-coverage articles report METR found Sol's detected reward-hacking rate the HIGHEST of any public model it has evaluated β primary METR post unfetched this run (403s), re-confirm before load-bearing use. Omniscience: minor accuracy gain over gpt-5.5 but hallucination rate HIGHER β NOT a council-honesty candidate (grok-4.20 unaffected). No gpt-5.5 deprecation announced (explicitly checked), but gpt-5.5 is now price-dominated: Sol matches its $5/$30 with better agentic numbers; Terra β beats it at half price. Willison's hands-on: Sol competent but did not outperform Fable-5 on his complex coding tasks.
β¦grok-4.5 (SpaceXAI, GA 2026-07-08 devs / 07-09 public): the provider rebranded xAI β SpaceXAI at this launch (affects all xAI-provider rows + the access matrix β confirm the XAI_API_KEY/base-URL path is unchanged before relying on it). Pricing $2/$6, cached input $0.50; ~80 TPS. Built on the 1.5T-param "V9" foundation, trained alongside Cursor (available in Cursor all-plans + Grok Build + xAI console + OpenRouter). Benchmarks are AA-verified (Intelligence Index 54 #4, Coding Agent Index 76, GDPval-AA v2 1543 Elo #4) plus launch-coverage SWE numbers β RECORD BOTH sides of two disagreements: (1) DeepSWE provider-harness 62.0% vs neutral-harness 53%; (2) SWE-bench Pro lists Opus 4.8 at 69.2% (max) in xAI's chart vs this doc's roster 64.3% Pro β different harness/effort, don't overwrite. Context window SHRANK to 500K (grok-4.3 = 1M, grok-4.20 = 2M) β grok-4.5 is NOT a long-context pick. NO AA-Omniscience component published β not a council honesty seat (grok-4.20 keeps that slot). Modalities: OpenRouter lists text in/out; image/video NOT confirmed this run β do not route OCR/PDF here until confirmed. Not in the EU at launch (xAI expects mid-July). A 2026-06-29 "private beta only, no public access" report was superseded by the 07-08 public GA β pre-GA status flips fast (Fable-5 lesson), verify at GA.
βgrok-4.6 (SpaceXAI, GA 2026-08-12, OpenRouter x-ai/grok-4.6): the direct successor to grok-4.5 β SAME $2/$6 base, SAME 500K ctx, knowledge cutoff Feb 1 2026, live day-one on OpenRouter/Vercel/Cloudflare + Cursor + Grok Build (double included usage in Cursor/Grok-Build for the first week). Pricing trap carried over from 4.5: prompts >200K tokens rebill the ENTIRE request at $4/$12 (a "fast" variant is 2Γ standard). OpenRouter weighted-avg input ~$0.74 because ~90.3% of tokens hit cache. Benchmarks: SpaceXAI's own β AA Intelligence Index 61, DeepSWE v1.1 65.9 (+11.9 gen-over-gen), APEX-Agents 57.5, CursorBench v3.2 69.9, FrontierCode v1.1 61.3. AA independent (direct-fetched the AA analysis article, NOT the model page β which was not fetched this run): Intelligence Index 61 = +5 vs grok-4.5, "matches GPT-5.6 Sol / Sol Max"; GDPval-AA v2 1753 Elo (behind only Opus 5, overlapping CIs with Fable-5 + Qwen3.8-Max); AA-Briefcase 1577 Elo (Fable-5-tier, behind the Opus 5 family); cost-per-task $0.84 (ties Kimi K3); efficiency profile ~53 turns / ~0.5B input tokens vs Opus 5's ~103 / ~2.0B. AA calls it strongest on agentic/knowledge-work + cost-efficiency, weaker on the coding rows engineering teams care about (DeepSWE 65.9 trails GPT-5.6 Sol Max ~73; no independent SWE-bench Pro/Verified anywhere). VentureBeat framing: "overtakes Kimi K3 and matches GPT-5.6 Sol for world's third best on AA." AA-INDEX SCALE DISCREPANCY (open): the AA analysis article implies Opus-5 = 63 while this doc's recorded v4.1 baseline is Opus-5 61 / Sol 59 β either AA rescaled the index again or the "61 = Sol" is a MAX-effort-tier comparison; not reconciled this run (needs a direct AA model-page fetch). Modalities: vision/tool-use NOT detailed in the announcement β do not route OCR/PDF here until confirmed. NO published AA-Omniscience component β not a council-honesty pick (grok-4.20 keeps that seat). Provider is SpaceXAI (xAI rebrand carried from the grok-4.5 launch). Sources fetched this run: x.ai/news/grok-4-6, artificialanalysis.ai/articles/grok-4-6-benchmarks-and-analysis, openrouter.ai/x-ai/grok-4.6, marktechpost.com (08-12), digitalapplied.com, eesel.ai; venturebeat.com via search.
βgrok-4.7 + muse-spark-1.3 (2026-09-21 refresh). grok-4.7 (SpaceXAI, GA 2026-09-21, x-ai/grok-4.7): direct successor to grok-4.6 β official page states ONLY $2/$6 (fast variant 2Γ price), availability (Cursor, Grok Build, Grok API, routers, clouds), and a vendor benchmark table; it does NOT state context, cutoff, cached price, modality, or whether the >200K-token $4/$12 whole-request rebill from 4.5/4.6 carries. The 500K ctx is from AA's model page. AA-independent (direct-fetched artificialanalysis.ai/models/grok-4-7): Intelligence Index 46 (xhigh) on v4.3.2, #16/202, cost-per-task $3.74, 38.8 TPS, "amongst the leading models in intelligence and reasonably priced" but "notably slow and very verbose" (240M output tokens across the eval). Do NOT compare $3.74 to grok-4.6's $0.84 β different index version, different task mix. Vendor table (x.ai/news/grok-4-7, SpaceXAI harness): CursorBench 4.0 46.3 / 4.6 40.4 / Sol 41.7 / Fable 5.1 51.8; DeepSWE v1.1 71.0@high / 65.2 / 72.7 / 70.0; EEBench 64.0 / 53.0 / 39.4 / 56.4; AA-Briefcase v1.1 1657 / 1546 / 1487 / 1678; Terminal-Bench 4.0 38.0 / 20.3 / 37.3 / 57.9; Harvey Legal Agent 19.6 / 15.8 / 2.5 / 6.7; HealthBench Professional 56.7 / 48.5 / 60.5 / 62.1; GDPval Elo 1695 / 1605 / β / 1735. Minor disagreement recorded: the vendor table lists grok-4.6 DeepSWE at 65.2 vs this doc's 65.9. No Omniscience component β not a council seat. No grok-4.6 EOL. muse-spark-1.3 (Meta, GA 2026-09-02, OpenRouter meta/muse-spark-1.3): multimodal reasoning model for long-running agentic/multi-agent/coding work; 1,048,576 ctx; text+image+video in. AA v4.3.2 Intelligence Index 48 (max), cost-per-task $1.60 β above GPT-5.6 Sol (47) and Grok 4.7 (46), one below Opus 5 high (48, tie). Pricing $1.25/$4.25 standard and $0.10/$0.20 "contributor" (Meta trains on prompts + completions, 100 RPM) is from SEARCH SNIPPETS (mindstudio.ai, openrouter listing, getdeploying) β not fetched at the model card this run; verify before any routing. Open weights "planned". No independent coding number; the AA index is the only third-party read. Sources fetched this run: x.ai/news/grok-4-7, artificialanalysis.ai/models/grok-4-7, artificialanalysis.ai/leaderboards/models, artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3, platform.claude.com pricing + models + deprecations, openrouter.ai/anthropic, cellcog.ai (opus-5.5 rumor forensics).
βkimi-k3 (Moonshot, GA 2026-07-16 API-first; 2.8T-param sparse MoE, largest open-weight model ever announced; weights PROMISED by 2026-07-27, license unspecified at launch β do NOT assume K2.x's Modified-MIT carries over): new Kimi Delta Attention (KDA) + Attention Residuals (AttnRes) architecture; ONE reasoning-effort level ('max') β no cheap-effort tiers; automatic context caching (no cache IDs/TTL), cache-hit $0.30/MTok, Moonshot claims >90% hit rates in coding workloads. AA-verified (direct-fetched 07-18): Intelligence Index v4.1 = 57, #4 of 187 per AA's page β launch coverage says #3; on known v4.1 values (Fable-5 60 / Sol 59 / Terra 55) a 57 slots #3, so AA's #4 implies an unlisted β₯57 entry (likely a gpt-5.5 high/xhigh effort-tier row β unresolved, both recorded); cost-per-task $0.94 (Sol $1.04, Opus 4.8 $1.80); 62 TPS (#91 β slow-ish); AA logged it verbose (130M output tokens across its eval) despite the vendor's "21% fewer output tokens than K2.6" claim. Arena.ai frontend-code #1, ahead of Fable-5 (independent). Vendor-only coding suite (KimiCode harness @ max effort, no third-party reproduction): Terminal-Bench 2.1 88.3 (Sol's 88.8 SOTA stands β not harness-comparable), DeepSWE 67.5, ProgramBench 77.8, FrontierSWE 81.2, SWE Marathon 42.0 (would crush grok-4.5's 29.0 IF independently reproduced). NO SWE-bench Pro/Verified anywhere. Moonshot's own framing: "mostly beats Opus 4.8 max and GPT-5.5 high; trails Fable-5 and Sol on aggregate." Willison hands-on: pelican test burned 13,241 reasoning tokens for 3,417 output tokens (~$0.25/query) β max-only effort has real small-task overhead; his cited "+732 Elo vs K2.6" is implausible as written (fetch artifact), the 1547 Elo itself is indicative (cf. grok-4.5's GDPval-AA 1543). Council seat evaluated β DECLINED (no Omniscience number, capacity-limited upstream, open-weights/price is not a council criterion). OpenRouter live day-one but single-provider with a standing capacity warning ("may return frequent 429 errors") until weights land. Sources fetched this run: openrouter.ai/moonshotai/kimi-k3, artificialanalysis.ai/models/kimi-k3, simonwillison.net (07-16), mlq.ai.
βqwen3.8-max (Alibaba/Qwen, GA 2026-08-03, previewed 07-19 at WAIC; 2.4T-param Sparse-MoE, ~95B active per token, native text+image+video in / text out): flagship Qwen model and the FIRST Max-class Qwen to be open-sourced (weights promised ~week of Aug-10; a Qwen3.8-27B ships open-weight simultaneously; license UNSPECIFIED at launch β do not assume prior Qwen terms carry). Context 1M nominal (max input ~991K / 983K with reasoning; max output 131,072; OpenRouter Cloud metadata lists 983,616 ctx). Pricing $2/$6 on DashScope/OpenRouter (cache-hit implicit $0.25, explicit-read $0.17); ONE aggregator (Qubrid AI) disagrees β list $2.87/$7.12, discounted $2.30/$5.69 on a 20% launch promo β treat $2/$6 as canonical, flag both. OpenRouter live day-one but single-provider (same capacity-risk shape as K3 at launch). Benchmarks: the ONLY independent axis pulled is AA Intelligence Index 53 β and it came from a search snippet, NOT a direct AA fetch (the model page 404'd this run), so treat as indicative-not-confirmed; on the v4.1 scale (Fable-5 60 / Opus-5 61 / Sol 59 / K3 57 / Terra 55 / Luna 51) a 53 lands mid-frontier and notably BELOW open-weight peer K3 (57). Everything else is Alibaba's own table (no independent SWE-bench reproduction): Terminal-Bench 2.1 86.6, SWE-bench Pro 67.7 (vs Fable-5 80), FrontierSWE 73.5, GPQA-D 92.6, PaperBench 93.0, IFBench 82.8; huge gen-over-gen jumps vs Qwen3.7-Max (DeepSWE 1.1 21.6β56.6, FrontierSWE 40.7β73.5, JobBench 31.3β53.4). Arena.ai #5 Text / #2 Vision β the one differentiated lane. AA flags it verbose (150M tokens generated during the eval vs ~63M median) and slow. China data-jurisdiction / state-law risk flagged in coverage (TechTimes: "QwenWork brings China state-law risk into enterprise workflows"; Bloomberg) β same non-sensitive-paths-only scoping vita applies to glm/kimi/deepseek; open weights (when they land) allow self-host. Positioning: direct price/capability competitor to grok-4.5 ($2/$6) and gpt-5.6-terra ($2.50/$15), open-weight rival to Kimi K3 β NOT a cheap-band (glm/kimi/minimax) entry. WATCH, zero routing. Sources fetched this run: marktechpost.com (08-03 + 07-19 preview), openrouter.ai/qwen/qwen3.8-max, alizila.com, technode.global; AA Index 53 via search snippet (artificialanalysis.ai/models/qwen3-8-max 404'd on direct fetch).
βglm-5.3 (Z.ai/Zhipu, GA 2026-08-14): a post-training-only refresh of glm-5.2 β SAME 743B base, every gain from extended post-training (Z.ai's own framing). Vendor-claimed strongest open-weights coder, +50% over glm-5.2; vendor-harness numbers DeepSWE 66.9, AutomationBench 48.2, GDPval-AA v2 1769 Elo β no independent SWE-bench Pro/Verified reproduction, same "vendor-only, re-confirm at AA" posture as glm-5.2/kimi-k2.7-code. The differentiated axis is CYBER: CyberGym 84.5% slightly surpasses Mythos 5 and GPT-5.6 Sol; working with China security teams Z.ai reports finding 2,436 vulnerabilities across 269 projects (some ~40yr-old); gaps remain vs closed frontier on deep exploitation (ExploitBench). Compatible with ZCode / Claude Code / OpenCode agents. Pricing NOT published as per-token β only GLM Coding Plan subscriptions surfaced (Lite $12.6 / Pro $56 / Max $117.6 per month); NOT on OpenRouter (glm-5.2 remains at $1.40/$4.40). API + OPEN WEIGHTS staged ~2 weeks out, after safety hardening β license UNSTATED (glm-5.2 was MIT; do not assume it carries). Context window unstated (glm-5.2 = 1M usable via IndexShare; likely same, same base β unconfirmed). AA Intelligence Index NOT published. Deliberately UNSCORED in the 1-10 grid (search-snippet-free AA + vendor-only coding = fabrication risk). WATCH, zero routing β glm-5.2 stays the strong-cheap-code pin. VITA-RELEVANT WATCH: the cyber lane is the first roster-adjacent path for authorized security-testing where opus-5/opus-4-8/fable-5 all refuse β revisit as a self-host pilot once weights + an independent CyberGym/ExploitBench number land. Sources fetched this run: the-decoder.com/zhipu-ai-releases-glm-5-3-β¦, explainx.ai/blog/glm-5-3-launch-β¦, unite.ai/z-ai-launches-glm-5-3-β¦, openrouter.ai/z-ai/glm-5.2 (5.3 not yet listed).
βgemini-3.7-flash (Google, GA 2026-08-13, google/gemini-3.7-flash): a cheap agentic/coding/document-QA workhorse, released only ~3 weeks after gemini-3.6-flash. Large gen-over-gen jumps (vendor): FrontierCode 1.1 Main 34.4β43.6, GDP.pdf 22β34, AutomationBench 17β30.4. Google claims it tops Claude on business workflows (vendor β no independent AA number pulled). Context 1,048,576 (1M), max output 65,536, multimodal, live on OpenRouter day-one. PRICING DISAGREEMENT (both recorded, verify before volume routing): vendor/news cite intro $0.75/$3.75 through 2026-12-31, reverting to $1.50/$7.50 on 2027-01-01; OpenRouter surfaced $0.375/$1.875 (exactly half the intro β possibly a batch/further-discount rate or a stale aggregator read). Knowledge cutoff not confirmed this run. Positioning: vita NEVER rostered 3.5/3.6 Flash β the budget-fast pin is still gemini-3.1-flash-lite (~$0.10/$0.40) β so 3.7-Flash is a genuinely NEW mid-tier sitting between flash-lite and gemini-3.1-pro ($2/$12); a candidate cheap-agentic/document challenger to the strong-cheap band, NOT a flash-lite replacement (5-8Γ the price). Deliberately UNSCORED in the 1-10 grid (no independent third-party benchmark). WATCH, zero routing until an independent (AA) Intelligence Index + a confirmed steady-state price exist. Sources fetched this run: pulse2.com/google-launches-gemini-3-7-flash-β¦, techtimes.com (324387 + 324366, 08-13), openrouter.ai/google/gemini-3.7-flash.
By-domain scoring (1β10, vita-calibrated)
Scores blend published benchmarks with fit for vita workloads. See Β§ Methodology.
| Domain | claude-opus-4-8 | gpt-5.5 | gemini-3.1-pro | kimi-k2.6 | v4-pro | glm-5.2 | sonnet-4-6 | minimax-m2.7 | v4-flash | grok-4.20 | grok-4.3 | fable-5 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Agentic coding (SWE-bench Pro / Verified) | 10 (64.3% Pro) | 9 (58.6%) | 8 (54.2%) | 9 (58.6%) | 9 (80.6% Verified, no Pro) | 9 (62.1% Pro β top open-weights) | 7 | 8 (56.2%) | 8 (~79% Verified) | 6 | ? (no SWE-bench published) | 10 (80.3% Pro β new leader) |
| Long-horizon task completion | 10 | 9 | 8 | 10 (300 subagents, 4k steps) | 8 | 8 (FrontierSWE/SWE-Marathon wins, high/xhigh thinking) | 7 | 7 | 7 | 6 | 7 (16-agent Heavy carryover, unverified) | 10 (built for minutes-long autonomous runs) |
| General reasoning (AA Intel Index) | 9 | 10 (57) | 10 (57) | 9 (54) | 9 (52) | 7 | 7 | 7 (50) | 7 (47) | 6 (48) | 8 (53) | 10 (60 on v4.1, #1 β confirmed 07-10) |
| Vision / screenshots / PDFs | 9 (2576px OCR) | 8 | 10 | 7 | 5 (text only) | 6 | 8 | 6 | 5 (text only) | 7 | 8 (text+image, video native β first frontier with native video) | 10 (high-res; GDP.pdf 29.8 top) |
| Long context (>200k) | 8 (1M beta) | 8 (400k) | 10 (1M+ stable) | 9 (1M) | 9 (1M, 83.5% MRCR) | 9 (1M usable, IndexShare) | 7 | 7 (200k) | 9 (1M) | 9 (2M) | 8 (1M API-exposed; 2M claimed; tiered pricing past 200k) | 9 (1M default) |
| Cost-efficiency @ quality | 5 | 5 | 8 | 9 | 9 (10 during May-5 promo) | 9 ($1.40/$4.40; glm-5.1 keeps the 10) | 8 | 9 | 10 | 3 | 9 | 3 ($10/$50 β priciest GA frontier) |
| Non-hallucination / honesty | 8 | 8 | 8 | 7 | 7 | 7 | 8 | 7 | 7 | 10 (78% Omniscience) | ? (component score not yet published β pilot decision blocker) | 8 (no Omniscience published) |
| Writing / voice fidelity | 10 | 9 | 8 | 7 | 6 | 7 | 9 | 7 | 6 | 7 | 6 (verbose by default β 88M tokens during AA eval) | 10 (clearer/warmer per launch notes) |
| Tool use stability | 10 | 9 | 8 | 9 | 8 | 9 (MCP-Atlas 77.0, near Opus 4.8) | 9 | 7 | 7 | 7 | 7 (assumed carryover, unverified) | 10 (SOTA agentic) |
| Streaming speed (TPS) | 7 | 7 | 9 | 9 | 4 (~34 TPS) | 7 (TTFT ~10.8s, thinking; async-friendly) | 9 | 10 | 8 | 7 | 10 (206.9 TPS, #1; TTFT 12.65s due to extended thinking) | 6 (long deliberate turns; minutes-long single requests) |
Blank cells = insufficient public data, scored conservatively by inference.
New 2026-07-05 entrants not yet in the grid (deliberate):
claude-sonnet-5,minimax-m3, andkimi-k2.7-codeare in the roster above but are NOT scored in this 1-10 table this refresh β none has verified third-party benchmarks yet (Sonnet 5: qualitative Anthropic claims only; M3: no AA numbers pulled; K2.7-code: Moonshot-proprietary suites only, no independent SWE-bench). Scoring them now would be fabrication. They're tracked as pilots/successors in the roster + routing sections; pull AA Intelligence Index + SWE-bench for all three next refresh, then add columns. (Same discipline the doc applied to glm-5.2's held general-reasoning score.)β AA Intelligence Index rescale (2026-07-10): artificialanalysis.ai moved to Index v4.1 with the gpt-5.6 launch. The parenthetical AA numbers in the grid above (gpt-5.5 = 57, kimi = 54, etc.) are OLD-index values and are NOT comparable to v4.1 values (Fable-5 60 / Sol 59 / Terra 55 / Luna 51). Do not mix scales when scoring; re-pull v4.1 values for the incumbent columns at the next full grid rebuild.
β AA rescaled AGAIN β Index v4.1.1 / "2026" (2026-09-03): the fable-5.1 leaderboard reads on a THIRD scale (AA labels it v4.1.1 / "Intelligence Index 2026"; components now GDPval-AA v2, ΟΒ³-Banking, Terminal-Bench v2.1, SciCode, HLE, GPQA-Diamond, CritPt, AA-Omniscience, AA-LCR). Current top values (do NOT mix with v4.1 or old-index): Fable 5.1 65.7 (#1) Β· Opus 5 63.0 Β· Fable 5 62.1 Β· Grok 4.6 60.9 Β· Kimi K3 59.7 Β· GLM-5.3 59.5 Β· GPT-5.6 Sol 58.9 Β· Gemini 3.8 Flash 58.7 Β· Qwen3.8-Max 58.1 Β· Opus 4.8 57.3 Β· GPT-5.5 56.3 (source: benchlm.ai mirror of AA, verified 09-02). This RESOLVES the 08-13 open flag (Opus-5 reads 63.0 β AA did rescale the index; the "61=Sol parity" was scale drift, not a max-effort artifact). Note the reshuffle: GPT-5.6 Sol (58.9) now sits BELOW Fable 5 and Opus 5 β no routing consequence (the OpenAI council seat is a separate gated proposal), but do not carry the old "Sol β Fable-5" framing.
gpt-5.6 family domain scores (new 2026-07-10 β SCORED: AA-verified third-party benchmarks, direct-fetched): Sol β agentic coding 9 (Coding Agent Index 80 #1, Terminal-Bench 2.1 88.8 SOTA; but SWE-bench Pro 64.6, ~15pts behind Fable-5); long-horizon 9 (Agents' Last Exam 53.6 top, OSWorld 62.6 with 85% fewer output tokens than Opus 4.8, native subagent spawning); general reasoning 10 (v4.1: 59, one below Fable-5); vision 8? (image input confirmed,
detail: originaloption new β no OCR benchmarks pulled); long context 9 (1M, MRCR 91.5); cost-eff 7 ($1.04/AA-task β 1/3 Fable-5, but $5/$30 sticker); honesty 7? (Omniscience: accuracy up vs 5.5 but hallucination rate UP β plus the METR reward-hacking caveat); writing/voice 8 (AA-Briefcase 2nd to Fable-5); tool use 9 (programmatic tool calling is genuinely new β unproven in vita's long-session shape); streaming ? (no TPS pulled). Terra β same shape one notch down (v4.1 55, Coding Agent 77.4, Terminal-Bench 87.4, MRCR 89.6) at half price β best value in the family. Luna β coding 8 (Terminal-Bench 84.7 > grok-4.5), reasoning 8 (v4.1 51), long context 3 (MRCR 41.3 β collapsed; hard disqualifier), cost-eff 8 ($0.21/AA-task β but 3-5Γ m2.7's sticker, ~20Γ v4-flash output; NOT a budget-band replacement). Fold into full grid columns at the next rebuild alongside the v4.1 re-pull.grok-4.5 domain scores (new 2026-07-09 β SCORED, unlike the 07-05 entrants, because it has AA-verified third-party benchmarks): Agentic coding 9 (SWE-bench Pro 64.7 β Opus 4.8; Terminal Bench 2.1 83.3 > Opus; Coding Agent Index 76). Long-horizon 9 (SWE Marathon 29.0 > Opus 4.8's 26.0; Cursor-trained agentic). General reasoning 9 (AA Intel Index 54, #4). Vision 7? (modality unconfirmed β OpenRouter lists text-only this run; do NOT route OCR/PDF here until confirmed). Long context 7 (500K usable β a REGRESSION vs the 1M tier). Cost-efficiency @ quality 9 (headline strength β $2/$6 + ~2Γ token efficiency β effective cost-per-task well under Opus/GPT-5.5). Non-hallucination ? (no Omniscience published β conservative, don't assume). Writing/voice 7 (not a voice model). Tool use 8 (native FC/JSON, Cursor-trained; not yet vita-stress-tested for long-session partial-failure shape). Streaming 7 (~80 TPS β mid; grok-4.3 is 207). Fold into a full grid column next refresh once numbers settle.
grok-4.6 domain scores (new 2026-08-13 β SCORED, AA-verified independent benchmarks like grok-4.5, unlike the WATCH-only qwen entry): agentic coding 8 (DeepSWE v1.1 65.9, +11.9 gen-over-gen, but TRAILS GPT-5.6 Sol Max ~73; CursorBench v3.2 69.9, FrontierCode v1.1 61.3 β all xAI-harness, no independent SWE-bench Pro/Verified β held a notch below Sol's 9); long-horizon 9 (GDPval-AA v2 1753 Elo, behind only Opus 5; AA-Briefcase 1577 Fable-5-tier; the ~53-turn/0.5B-input-token efficiency profile is the headline); general reasoning 9 (AA Intel Index 61, one below Opus-5, ties Sol per AA β caveat: gated on the index-scale-drift open question below; nominally +7 vs grok-4.5's old-index 54, do not treat as directly comparable); vision 7? (modality unconfirmed β announcement omits it; do NOT route OCR/PDF until confirmed); long context 7 (500K, unchanged from 4.5; >200K rebills at $4/$12 β NOT a long-context pick); cost-efficiency @ quality 10 (the headline strength β $2/$6 + $0.84/AA-task + ΒΌ the input tokens of Opus 5 β best effective cost-per-task at this intelligence tier); non-hallucination ? (no Omniscience published β conservative, grok-4.20 keeps the honesty seat); writing/voice 7 (not a voice model); tool use 8 (Cursor-trained agentic; not yet vita-stress-tested for long-session partial-failure shape); streaming ? (no TPS pulled this run). Fold into a full grid column next refresh once the AA-index scale is reconciled.
kimi-k3 domain scores (new 2026-07-18 β PARTIALLY scored: AA-verified reasoning/price only; coding deliberately unscored): general reasoning 9 (v4.1 = 57, #3-4 β above Terra 55, below Sol 59; first open-weight model at this tier); agentic coding ? (ALL SWE-style numbers are KimiCode-harness vendor claims β no independent SWE-bench Pro/Verified; scoring now would be fabrication, same discipline as the 07-05 entrants); long-horizon ? (SWE Marathon 42.0 vendor claim would be #1 by a wide margin β unverified); frontend/web code 10 (Arena.ai #1 ahead of Fable-5 β independent, and the one lane where K3 has a verified lead); long context 9? (1M nominal, no MRCR-style depth test pulled); vision 7? (text+image in per AA; zero OCR benchmarks); cost-efficiency 7 ($0.94/AA-task is frontier-cheap, but $15/MTok output + max-only reasoning effort = real overhead on small tasks β ~$0.25 for a trivial SVG); honesty ? (no Omniscience); writing/voice ? (untested); streaming 5 (62 TPS #91, plus upstream 429s). Full grid column only after weights land + at least one harness-comparable coding number exists.
claude-opus-5 domain scores (new 2026-07-25 β SCORED: AA-verified third-party benchmarks, direct-fetched): general reasoning 10 (AA Intelligence Index v4.1 = 61, narrowly #1 of all models β one above Fable-5's 60); agentic coding 10 (joint-first AA Coding Agent Index @ xhigh in Claude Code, highest SWE-Atlas-QnA β BUT no independent SWE-bench Pro number pulled; the "10" rests on AA Coding-Agent + GDPval, flag if a Pro number lands lower); long-horizon 10 (record GDPval-AA v2 1861 + AA-Briefcase 1720; OSWorld-2.0 beats Fable-5 at ~1/3 cost β the successor to opus-4-8's 10); vision/OCR 9? (image input confirmed, no OCR benchmark pulled β held at opus-4-8's 9); long context 8 (1M nominal but secondary-source + no MRCR pulled β held at opus-4-8's 8 pending model-card confirmation); cost-efficiency 7 (top intelligence at Opus $5/$25 AND ~26% cheaper-per-task than Fable-5 β the new value-frontier; up from opus-4-8's 5, but still Opus-tier sticker so not a cheap-band pick); non-hallucination 8? (no Omniscience pulled; lowest behavioral-audit score of any recent Claude = best-behaved, but that is NOT an Omniscience number β held at 8); writing/voice 10 (Opus lineage, "thoughtful/proactive" per launch); tool use 10 (SOTA agentic, joint-#1 coding agent); streaming 7? (no TPS pulled; 2.5Γ fast-mode available). Fold into a full grid column next refresh alongside an independent SWE-bench-Pro number + AA Omniscience + TPS.
qwen3.8-max β deliberately UNSCORED (2026-08-05): the only independent number pulled is AA Intelligence Index 53, and it's a SEARCH SNIPPET (AA model page 404'd on direct fetch β not model-card-confirmed). All coding numbers (SWE-bench Pro 67.7, Terminal-Bench 86.6, FrontierSWE 73.5) are Alibaba's own table with no independent SWE-bench reproduction; vision rests on an Arena leaderboard rank (#2), not an OCR/PDF benchmark. Scoring any 1-10 cell now would be fabrication β same discipline the doc applied to the 07-05 entrants (sonnet-5/m3/k2.7-code) and to K3's coding rows. Provisional-only reads for the next-refresh scoring pass: general reasoning ~7-8 (AA 53 IF confirmed β below K3's 9-at-57), vision potentially high (the differentiated lane β gate on an AA vision/OCR number), cost-efficiency mid (same $2/$6 as grok-4.5 but verbose+slow erodes effective cost-per-task). Add real columns only after (a) a directly-confirmed AA index, (b) one independent coding number, and (c) an AA vision/OCR number exist.
glm-5.3 + gemini-3.7-flash β deliberately UNSCORED (2026-08-14): neither has ANY independent third-party benchmark. glm-5.3's coding/cyber numbers (DeepSWE 66.9, CyberGym 84.5, AutomationBench 48.2, GDPval-AA v2 1769) are all Z.ai's own harness with no AA Intelligence Index published; gemini-3.7-flash's numbers (FrontierCode 43.6, GDP.pdf 34, AutomationBench 30.4, "tops Claude on business workflows") are all Google's own. Scoring any 1-10 cell now would be fabrication β same discipline the doc applied to qwen3.8-max, kimi-k3's coding rows, and the 07-05 entrants. Add real columns only after an independent (AA) Intelligence Index β and for glm-5.3, an independent CyberGym/ExploitBench reproduction for the cyber lane β exists.
claude-fable-5-1 domain scores (new 2026-09-03 β SCORED: has an AA-independent Intelligence Index, unlike the WATCH entries): general reasoning 10 (AA Index 65.7 on the v4.1.1/2026 scale, new #1 β above Opus 5 63.0); agentic coding 10 (Terminal-Bench-Science 52.6 near-2Γ Opus 5's 29.0; CursorBench 73.4 β but these + Terminal-Bench 4.0 55.8 are Anthropic's own harness, no independent SWE-bench Pro pulled this run; the 10 rests on the AA index + GDPval, flag if an independent Pro number lands lower); long-horizon 10 (GDPval-AA v2 1853, top; AutomationBench 17.1β31.4; OSWorld 2.0 41.7 > fable-5's 36.1); vision/OCR 10? (image input; no OCR benchmark pulled β held at fable-5's 10); long context 9 (1M / 128K out β held at fable-5's 9); cost-efficiency 4 (still $10/$50 sticker β priciest GA frontier β but the 75% cache-read cut + token-efficiency lift it above fable-5's 3); non-hallucination 8? (AA-Omniscience is now an index COMPONENT but the standalone number wasn't isolated this run β held at 8); writing/voice 10 (Fable lineage); tool use 10 (SOTA agentic). Fold into a full grid column next refresh alongside an independent SWE-bench-Pro number + an isolated AA-Omniscience read. The only thing that changes vs fable-5's column is cost-efficiency 3β4 and reasoning 10-at-60 β 10-at-65.7 β it's a strict upgrade in the same niche, not a new lane.
New AA-leaderboard sightings (2026-09-03 β NOT scored, leaderboard-only): the AA v4.1.1 list surfaced five entries with an index number but no announcement/pricing/context fetched this run β Gemini 3.8 Flash (58.7), GLM-5.3-Flash (57.5), Meta Muse Spark 1.2 (56.8 β first Meta model near the top; Meta has never been on vita's roster), Qwen3.8-Flash-Next (55.8), and glm-5.3 now with a published AA index (59.5) where it was unpublished at the 08-14 refresh. All WATCH / zero-routing until an announcement + pricing + (for the cheap-band candidates) an independent benchmark land. Muse Spark is the one worth an explicit look next refresh β a new-vendor top-15 entrant is a bigger signal than another Flash SKU.
β AA rescaled a FOURTH time β Index v4.3 (published 2026-09-07) / v4.3.2 (recorded here 2026-09-21): Terminal-Bench 2.1 β 4.0 (harder terminal tasks across software/ML/science/ops/security/hardware/media), ΟΒ³-Banking REPLACED by AutomationBench-AA (Zapier, 657-task private set); private test sets now 45% of weight. Components: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, HLE, GDP.pdf, CritPt, AA-Omniscience, AA-LCR v1.1. Every v4.1.1 number in the blocks above (Fable 5.1 65.7 / Opus 5 63.0 / Astra 61 / Sol 58.9 / Grok 4.6 60.9 / K3 59.7 / GLM-5.3 59.5) is on a RETIRED scale β do not mix. v4.3.2 top values (direct-fetched leaderboard, effort tier in parens): Fable 5.1 53 (max, xhigh) Β· GPT-6 Astra 53 (max) Β· Astra 52 (xhigh) Β· Fable 5.1 51 (high) Β· Astra 51 (high) Β· Opus 5 51 (max) Β· Opus 5 50 (xhigh) Β· Astra 50 (medium) Β· Fable 5.1 49 (medium) Β· Opus 5 48 (high) Β· Muse Spark 1.3 48 (max) Β· GPT-5.6 Sol 47 (max) Β· Fable 5.1 47 (low) Β· Grok 4.7 46 (xhigh). Cost-per-task on the same board: Fable 5.1 max $7.63 / xhigh $5.98 / high $3.91, Opus 5 max $5.86 / xhigh $4.88 / high $3.61, Astra max $3.26 / xhigh $2.31 / high $1.73, Sol max $1.99, Muse 1.3 $1.60, Grok 4.7 $3.74. Two ordering changes vs the 09-04 read: (a) Astra now TIES Fable 5.1 at the top instead of trailing by 5 β the "priced like a leader without being one" line softens to "a leader at ~43% of Fable 5.1's cost-per-task", though its 51% hallucination + no-OpenRouter + non-ZDR-unknown caveats all stand; (b) Sol (47) drops below Meta's Muse Spark 1.3 (48). General-reasoning cells in the mini-blocks are held as-is this run (relative order among the Anthropic/OpenAI leaders is unchanged); re-score the column at the next full grid rebuild on v4.3.2 only.
grok-4.7 domain scores (new 2026-09-21 β SCORED where AA-independent, like grok-4.6): general reasoning 8 (AA v4.3.2 = 46 xhigh, #16 β one below Sol's 47 and two below Muse 1.3; on the same scale Opus 5 max is 51 β held a notch below grok-4.6's 9 because the new scale places it further from the leaders than the old one did); agentic coding 8 (vendor-harness only: CursorBench 4.0 46.3 > Sol 41.7, DeepSWE 71.0 β Fable 5.1 70.0 but @high effort, Terminal-Bench 4.0 38.0 β Sol 37.3 β no independent SWE-bench Pro/Verified; same discipline as 4.6); long-horizon 8 (vendor GDPval 1695 / AA-Briefcase 1657 β both BELOW Fable 5.1; no independent GDPval-AA pulled this run, so down from 4.6's AA-backed 9); vision ? (modality unstated in the announcement β same standing caveat as 4.5/4.6, do NOT route OCR/PDF); long context 7? (500K per AA; >200K rebill unconfirmed for 4.7); cost-efficiency 8 ($2/$6 sticker unchanged, but AA's $3.74/task + "very verbose" flag mean the effective-cost headline that earned 4.6 a 10 has NOT been demonstrated for 4.7 β do not inherit the 10); non-hallucination ? (no Omniscience component); writing/voice 7 (not a voice model); tool use 8 (Cursor-trained lineage, unverified in vita's shape); streaming 4 (38.8 TPS per AA β slow). Fold into a full grid column at the v4.3.2 rebuild.
muse-spark-1.3 β reasoning-only score (new 2026-09-21): general reasoning 8 (AA v4.3.2 = 48 max β the ONE independent number, direct-fetched from the leaderboard; ties Opus 5 high, beats Sol); cost-efficiency 9? ($1.60/task is the cheapest top-12 entry, but the $1.25/$4.25 sticker is search-snippet-only); everything else ? (no coding/vision/honesty/TPS numbers pulled; data-training "contributor" tier is a hard NO for vita). WATCH; full column only after model-card pricing + one independent coding number.
gpt-6-astra domain scores (new 2026-09-04 β SCORED: has an AA-independent Intelligence Index + Coding Agent Index, unlike the WATCH-only qwen/glm-5.3/gemini-3.7-flash entries): general reasoning 9 (AA Index 61 max β TIES Sol, below Fable 5.1's 66 and Opus 5's 63 on the same v4.1.1 scale; held one below the 10s); agentic coding 9 (AA Coding Agent Index 67 β Opus 5 β an INDEPENDENT number, stronger than grok-4.6 had; the vendor-harness Terminal-Bench 4.0 57.7 / DeepSWE 74.1 are recorded but not scored on); computer-use / long-horizon 10 (OSWorld 2.0 72.6 at ~2Γ Sol's speed is the differentiated lane + the genuine gen-over-gen leap β vendor-run Responses-API harness, so flag if an independent number lands lower; Agents' Last Exam 59.3); vision/OCR ? (image input presumed but no OCR/document-VQA number pulled β do NOT route vision here until confirmed); long context 8? (1.05M nominal / 128K out, no MRCR-style depth test pulled β held conservative, and note Sol-family Luna's MRCR collapse means "nominal 1M" is not a depth guarantee for this vendor); cost-efficiency 3 ($10/$50 sticker β 2Γ Opus 5 / 2Γ Sol for an AA-index TIE, and AA measured it 75% more expensive per task than Sol at max effort; worse value than Fable 5.1's 4, which at least holds the #1 index); non-hallucination ? (no isolated AA-Omniscience pulled; AA logged 51% hallucination at max effort β high β so conservatively NOT a strength; grok-4.20 keeps the council honesty seat); writing/voice ? (not researched β no voice/style benchmark pulled; do not route voice-critical paths here); tool use 9 (programmatic tool calling carryover from the 5.6 Responses API + strong agentic/computer-use numbers β unproven in vita's long-session shape); streaming ? (no TPS pulled). Fold into a full grid column next refresh alongside an independent SWE-bench-Pro number, an AA-Omniscience read, a vision/OCR number, and TPS. The one-line read: strong on computer-use + gated cyber, a TIE on measured general intelligence, priced like a leader without being one β WATCH, not a routing swap.
claude-opus-5-5 domain scores (new 2026-09-23 β SCORED: AA-independent index + component leads, direct-fetched): general reasoning 10 (AA v4.3.2 = 58 max, #1 of 212, a 5-point lead over Fable 5.1 and Astra at 53 β the largest gap any model has held on this scale; takes the 10 opus-5 held); agentic coding 10 (Terminal-Bench 4.0 66.4 and CursorBench 4.0 57.8 are Anthropic-harness, but AA's independent run reads 59.6 on Terminal-Bench 4.0 β a TIE with Astra, not a rout; the 10 rests on the AA composite + the GDPval/AA-Briefcase leads, flag if an independent SWE-bench Pro lands lower); long-horizon 10 (GDPval-AA v2.1 1846 +111 vs Fable 5.1, AA-Briefcase 1822 +143 β both independent, both the largest margins on the board; leads AutomationBench-AA); vision/OCR 9? (multimodal confirmed, Chartography 89.0 vendor β no independent OCR/document-VQA number; held at opus-5's 9); long context 8? (1M model-card-confirmed this time, unlike opus-5's secondary-source 1M β but AA has it BEHIND on AA-LCR, its long-context-reasoning component, so do NOT raise this above opus-5's 8 and do not treat 1M nominal as depth); cost-efficiency 8 (up from opus-5's 7: $4/$20 is the cheapest frontier-leader sticker ever rostered and the 0.05Γ cache read is real β but capped at 8, NOT 9-10, because AA's cost-per-task is FLAT at $5.98 vs opus-5's $5.86 on 1.6Γ verbosity; gpt-6-sol does the same work for ~1/6 the per-task cost at 10 fewer index points, so this is not the value frontier); non-hallucination / honesty 9 (first real move on this row in the doc's history β it LEADS AA-Omniscience, the component that has gated every council-honesty challenger since grok-4.20; held at 9 not 10 because the isolated Omniscience number vs grok-4.20's 78% was not pulled this run, and AA's own framing is that Anthropic's Omniscience lead is calibration-driven while xAI's position is accuracy-driven β different angles, see the council re-examination in Open Questions); writing/voice 10 (Opus lineage); tool use 10 (SOTA agentic; +the lowest prompt-injection success rate Gray Swan has recorded, tied with Fable 5.1 β independent); streaming 8 (+30% output speed vs opus-5 per Anthropic; no TPS number pulled β up from opus-5's 7? on the vendor claim only). The one-line read: a strict upgrade on every axis that matters AND 20% cheaper per token β but the per-TASK cost is flat, so adopt it for the quality, not for the savings.
claude-sonnet-5-5 domain scores (new 2026-09-29 β SCORED: AA-independent index, direct-fetched): general reasoning 9 (AA v4.3.2 = 56 max, #3 of 216 β two below opus-5-5's 58, three above Fable 5.1/Astra at 53; a +18 jump from sonnet-5's 38 at the same price, but held at 9 not 10 because opus-5-5 keeps a clear lead and opus-5-5 at xhigh scores the identical 56 for less than half the cost-per-task); agentic coding 9 (vendor Terminal-Bench 4.0 70.6 BEATS opus-5-5's 66.4 and CursorBench 4.0 55.5 β 57.8 β genuinely strong, but Anthropic-harness only, and the 09-23 precedent is that AA's independent Terminal-Bench run read ~7 points below Anthropic's for opus-5-5; no independent SWE-bench Pro pulled, so held one below the 10); long-horizon 9 (GDPval-AA v2.1 1844 and AA-Briefcase 1811 are within 2 and 11 Elo of opus-5-5's 1846/1822 β near-parity at half the per-token price, vendor-reported); vision/OCR 9? (OSWorld 2.1 80.1 β opus-5-5's 81.8, Chartography 61.6 vs 64.4; "first Sonnet to beat PokΓ©mon Red from screenshots only" is a real multimodal-planning signal but not a benchmark β no independent OCR/document-VQA number, held at opus-5-5's 9 minus nothing); long context 8? (1M model-card-confirmed / 128K out, but no MRCR or AA-LCR depth number pulled β same discipline as opus-5-5, nominal 1M is not depth); cost-efficiency 7 (the row that will mislead if read fast: $2/$10 is half opus-5-5's sticker and the cache shape is standard 0.1Γ, but AA's cost-per-task is $7.60 β the HIGHEST of any model in the top 5, above opus-5-5's $5.98 and 49% above sonnet-5's $5.09 β so it is scored BELOW opus-5-5's 8, not above it, despite the cheaper sticker; Anthropic's "up to 30% less per task" on real workloads is recorded but unreproduced independently); non-hallucination ? (no isolated AA-Omniscience read pulled and it does not lead the component β opus-5-5 does; conservatively not a strength, grok-4.20 keeps the council honesty seat); writing/voice 8? (Sonnet lineage + "polished documents, slides, spreadsheets" + "a sharp eye for design" are launch prose, not a voice benchmark β explicitly NOT scored 9-10, because vita's voice-critical paths are the one place a wrong call is hardest to detect; the 26 sonnet-4-6 refs stay held pending a drift check); tool use 9 (strong agentic numbers, adaptive thinking, but unproven in vita's long-session shape); streaming 9 (137.6 TPS per AA, "well above average" and the fastest Sonnet to date β one of only two rows where it clearly beats opus-5-5). The one-line read: a very large capability jump for exactly zero per-token cost, whose per-TASK cost is disputed in opposite directions by vendor and independent β adopt it on the bounded pilot modes, not across voice paths, and watch tokens.
claude-haiku-5-5 β deliberately UNSCORED (2026-09-29): announced, not shipped. No ID, price, context window or benchmark exists at any primary source; the only fact is Anthropic's own "coming weeks" commitment on the Sonnet 5.5 page. Scoring any cell β or carrying haiku-4-5's numbers forward β would be fabrication. No provisional reads offered: unlike the WATCH entries (qwen/glm-5.3/Luna), which at least have a vendor table to be skeptical of, this one has nothing at all.
gemini-4-argon β PARTIALLY scored (new 2026-10-02: AA-independent index + four isolated components, direct-fetched; the rest deliberately unscored): general reasoning 9 (AA v4.3.2 = 53 high, #8 of 224 β ties Astra and Fable 5.1 max, five below opus-5-5's 58 and three below sonnet-5-5's 56; a +23 jump over gemini-3.1-pro-preview's 30, the largest gen-over-gen move this doc has recorded for any vendor, which is why it earns a 9 rather than the 8 the raw rank would suggest); long-horizon / agentic 10 (#1 on AutomationBench-AA at 78%, +7 over sonnet-5-5 max β independent, and the one row where it leads the entire board; AA-Briefcase 1494 Elo on a 65% rubric pass rate, the highest ever recorded); agentic coding 7 (AA's Terminal-Bench 4.0 = 57, FOURTH behind sonnet-5-5 64 / opus-5-5 60 / Astra 59; vendor DeepSWE 77.9 and CWE-bench 68 are Google-harness; Code Arena ranks it #8 for web dev, a real gap, so this is scored BELOW the Claude leaders despite the vendor framing); vision / video 9? (LVBench 91.7 long-video SOTA and "professional chart analysis" are the differentiated lane and genuinely new β but both are Google's own numbers with no independent OCR/document-VQA read, so held at a question mark, NOT raised to gemini-3.1-pro's 10); long context 8? (1M ctx confirmed by AA; the 1M OUTPUT limit is the real headline and has no analogue on the roster β but no MRCR or AA-LCR depth number exists, and AA-LCR is a component it does not lead, so nominal is not depth); cost-efficiency 5 (the row most likely to be misread: $1.99/task looks like the best value in the top 10, but the $2/$10 is an UNDATED PROMOTION reverting to $4/$20 β $3.98/task, ~22% ABOVE Astra's $3.26, and it burns 62k output tokens/task vs Astra's 27k for equal quality; scored on the price that will hold, not the one on the page); non-hallucination / honesty 8 (15% AA-Omniscience hallucination, the lowest of any model scoring 45+ β genuinely best-in-class and a real result β but held at 8, NOT 9-10, because accuracy is only 50% vs Astra's 63% and the-decoder reads the composite knowledge score at 42 vs Astra 43: it abstains more rather than knowing more. A calibration win is not a knowledge win, and this is why it does NOT challenge grok-4.20's honesty seat); cyber ? (the "autonomously find, validate and patch vulnerabilities" claim + shipping without cyber guardrails to Fairwind defenders is the most aggressive posture any vendor has taken, and CWE-bench v1 68 ties first β but it is vendor-only AND unreachable, so unscored; if it ever reaches GA this becomes the strongest authorized-security-testing candidate on the roster, ahead of Fable 5.1's discover-not-exploit path and glm-5.3's unverified CyberGym); writing/voice ? (Arena.ai text #1 at 1525 is a crowd-preference signal, not a voice-fidelity benchmark β do NOT route voice paths on it); tool use ?; streaming ? (AA lists TPS as N/A). The one-line read: a genuine return to the frontier with one clear independent lead (agentic automation) and one genuine first (1M output), priced on a promotion, trailing the Claude leaders on coding and composite intelligence β and ALL of it academic until an API id exists. Full column only after GA + an API id + a confirmed steady-state price.
gpt-6.1-sol β PARTIALLY scored (new 2026-10-02: AA index + cost-per-task across four effort tiers are independent; every capability number is OpenAI's own harness): general reasoning 9 (AA v4.3.2 = 52 max, up from gpt-6-sol's 48 β one above Astra xhigh, ties Astra xhigh/sonnet-5-5 xhigh, six below opus-5-5's 58; held at 9 because it is a near-frontier score, not a frontier-leading one); cost-efficiency 10 (the headline, and the first 10 awarded on this row to anything but a budget-band model: 52 @ $0.72/task, 51 @ $0.39, 50 @ $0.32, 48 @ $0.21 β its high tier is cheaper per task than any tier of any model in AA's top 25, and it delivers ~90% of opus-5-5's index score at ~12% of its per-task cost; the $0.10 cached input and the gentle $4/$15 long-context rebill both compound this); agentic coding ? (DeepSWE +6.4 over gpt-6-sol and "matches Astra" is OpenAI's harness with no independent SWE-bench Pro/Verified and no AA Coding Agent Index pulled β scoring it would be fabrication, same discipline applied to qwen3.8-max and K3's coding rows); long-horizon / computer-use ? (AutomationBench +2.2 over opus-5-5 at ~β cost and OSWorld 2.0 within 2.1 of Astra at ~1/7 cost/task are the two most interesting claims on the model and both are vendor-run β these are exactly the numbers to demand independently before any agentic routing); long context 3? (1.1M nominal / 128K out, but gpt-5.6-luna's MRCR 41.3 collapse has never been re-tested anywhere in the gpt-6 line β carried forward as a hard caveat rather than a clean slate, so the nominal number earns no credit until an MRCR read exists); non-hallucination ? (vendor reports low-reasoning factual errors 11.4% β 7.7%, unreported-broken-tool 4.9% β 2.1%, and safety-restriction-bypass attempts 64.4% β 23.5% β a large self-reported improvement with no AA-Omniscience component, so not a strength and not a council-honesty candidate); vision/OCR ?; writing/voice ?; tool use ?; streaming ? (no TPS pulled). The one-line read: on the ONE axis that is independently measured it is the best value on the board by a wide margin, and on every capability axis we are taking OpenAI's word for it β which is precisely the right shape for a low-volume council seat and the wrong shape for a volume routing swap. Full column after an independent coding number + an MRCR re-test + the vendor announcement actually fetched (openai.com 403'd this run).
gpt-6-sol / gpt-6-luna / deepseek-v4.1-flash β deliberately UNSCORED (2026-09-23): Sol has exactly ONE independent number (AA v4.3.2 = 48 max, $0.13β$1.06/task) and no component breakdown, no coding number, no Omniscience read; Luna and V4.1-Flash have no independent benchmark at all and snippet-only pricing. Scoring 1-10 cells now would be fabrication β same discipline applied to qwen3.8-max, glm-5.3, gemini-3.7-flash and the 07-05 entrants. Provisional reads for the next pass, not scores: Sol's lane is cost-per-task at slightly-below-frontier intelligence (48 vs Opus 5's 51 at ~1/6 the per-task cost β the value seat, which is exactly why it retargets the council OpenAI proposal); Luna's lane is the budget band but is gated on an MRCR re-test (its 5.6-era predecessor collapsed at 41.3); V4.1-Flash is gated on modelling the peak/off-peak price split against vita's actual job clocks. Add real columns only after an AA index (Luna, V4.1-Flash), one independent coding number (Sol), and model-card pricing confirmation (all three).
Current vita routing (.aloop/config.json)
Audit as of 2026-04-23. Each mode pins a model for a task class; unpinned modes fall to the default (currently Claude via scripts/inference/claude_backend.py).
| Mode | Current model | Task type | Status |
|---|---|---|---|
silicon-zack / silicon-zack-chat |
silicon-zack (Modal) |
Identity/voice fine-tune | β correct β custom weights |
epd47_generate (.aloop/config.json:39), media_deep_dive (:90) |
anthropic/claude-sonnet-5 |
Creative + analytical writing | πΆ RE-ASSERTED 2026-10-02 (rec 4) β proposed 09-29, NOT applied (verified: both still read anthropic/claude-sonnet-5). No new argument is offered; the 09-29 one below stands unchanged. Carried as a batch with research-proposal:41 and the costs.py row. Prior (09-29): BUMP TO claude-sonnet-5-5 PROPOSED β same $2/$10, AA v4.3.2 38 β 56 (#58 β #3). The pilot started 09-21 and has run 8 days clean; 5.5 is a strict same-price upgrade over the model being piloted, so the right move is to re-point the pilot, not widen it. Non-voice analytical modes ONLY β every voice/composer path stays on sonnet-4-6 until a drift check reads clean. Watch TOKENS, not just quality: AA measures 5.5 at $7.60/task vs sonnet-5's $5.09, so if the verbosity read is the one that holds here, these two modes get more expensive at identical per-token pricing. data/tokens/YYYY-MM.json is the instrument. |
epd47_explore, research_analyze, research_extract, research_synthesize, background, deep_analysis, synthesis, speculative, research |
deepseek/deepseek-v4-flash |
Research/synthesis | β swapped from M2.7 (V4-Flash bake-off winner) |
daily_rollup, media_score, media_summarize, monitoring, trend_analysis, view_rebuild, micro_consolidation, rss_process |
google/gemini-3.1-flash-lite-preview |
High-volume triage | β correct β cheap, fast, good-enough |
gemini-flash-test |
google/gemini-3.1-flash-lite-preview |
Test mode | β correct |
sleep_build |
anthropic/claude-opus-4-7 |
Autonomous sleep build phase | π DELIBERATE HOLD β auto-marked 2026-07-18 per the 07-10 self-terminating rule. The opus-4-7β4-8 bump was recommended at 6 consecutive reviews (06-17 β 07-10) and never applied; silence = declined by inaction. No longer re-surfaced in review emails. CORRECTED 2026-10-02: the trio is now a PAIR β scripts/reports/generate_html_v2.py NO LONGER EXISTS (only stale .pyc under __pycache__/), so the live companion pin is just scripts/cadence/daemon.py:258. DELIBERATE_HOLDS already listed only those two; this doc was the stale copy. The quarterly reaffirm window opens THIS MONTH (~Oct 2026) β eventual target is now opus-5-5, not opus-5 or opus-4-8. Still not re-surfaced; one tap overturns. |
(unpinned: executive_reflect, sleep, coaching, doctor, voice, telegram, etc.) |
fall-through Claude | Real-time + high-stakes | β correct β Claude is the right default |
council seats (flows/council/setup_models.py:5-9) |
opus-5-5 / gpt-6-sol / gemini-3.1-pro-preview / grok-4.20 / deepseek-v4-pro | Multi-model deliberation | πΆ OpenAI SEAT RETARGET PROPOSED β gpt-6-sol β gpt-6.1-sol (rec 3, NEW). The 09-23 proposal to move this seat off gpt-5.5 landed on 09-29; seven days later OpenAI superseded the target at the identical $2/$10. AA v4.3.2 48 β 52, cost/task $1.05 β $0.72. One line, same price, strictly better measured intelligence. This is the second retarget in ten days, which is a statement about OpenAI's release cadence rather than about this doc's judgement β but leaving the seat on a model superseded one week after adoption is the kind of drift this flow exists to catch. Claude seat UNCHANGED (opus-5-5 keeps it). Argon DECLINED for a seat β 53 < opus-5-5's 58, and its Omniscience "lead" is calibration at 50% accuracy, so it challenges neither opus-5-5's intelligence seat nor grok-4.20's honesty seat. |
simple-council, council-review, deep-council, research/codebase-research, vision-strategy, fusion-council flows |
opus-5-5 / gpt-6-sol / gemini-3.1-pro / grok-4.20 | Peer-model fan-out | β
APPLIED β verified 2026-09-29, closing a drift this doc had promised since 07-25 and carried unapplied for nine weeks. All six now read anthropic/claude-opus-5-5 + openai/gpt-6-sol: simple-council/FLOW.yaml:40, council-review/FLOW.yaml:38, deep-council/FLOW.yaml:28, vision-strategy/FLOW.yaml:169, research/codebase-research/FLOW.yaml:172, fusion-council/setup_models.py:16. All eight council entry points now agree for the first time in this doc's history. |
scripts/council/core.py COUNCIL_DEFAULT_MODELS |
opus-5-5 / gpt-6-sol / gemini-3.1-pro / grok-4.7 / kimi-k3 | Programmatic council | πΆ SAME RETARGET APPLIES HERE β :23 gpt-6-sol β gpt-6.1-sol (rec 3). Both council entry points must move together or the eight-way agreement achieved on 09-29 breaks on its first week. :22 + :33 (the "claude" override) stay on opus-5-5. Council roster recommendation this run: ONE SEAT CHANGES (OpenAI), everything else KEEPS. gemini-4-argon declined (unroutable β no API id β and 53 < 58 even if it were reachable). sonnet-5-5 still declined per 09-29. grok-4.20 keeps the honesty seat, and argon's 15%-hallucination result does NOT dislodge it: 50% accuracy vs grok-4.20's accuracy-driven position means they are strong on different halves of the same metric. |
flows/research-proposal/FLOW.yaml:41 |
gpt-4.1-mini / gemini-2.5-flash-preview-05-20 / deepseek-r1 / claude-sonnet-5 |
4-vendor proposal panel | πΆ THREE OF FOUR ARMS ARE STALE, not one (rec 6, NEW 2026-10-02). The 09-29 review proposed bumping only the Anthropic arm and never read the rest of the line. Full current value: ["openai/gpt-4.1-mini", "google/gemini-2.5-flash-preview-05-20", "deepseek/deepseek-r1", "anthropic/claude-sonnet-5"]. gpt-4.1-mini is NOT deprecated (verified at OpenAI today β see rec 5) but is two generations behind gpt-6-luna at 10Γ the price; gemini-2.5-flash-preview-05-20 is a preview id three generations back; deepseek-r1 predates the entire v4 line. A panel whose job is vendor DIVERSITY is currently comparing four models nobody would pick. Proposed: openai/gpt-6-luna + google/gemini-3.1-flash-lite + deepseek/deepseek-v4-flash + anthropic/claude-sonnet-5-5. One line, four ids, and it is the cheapest fix on this list. |
scripts/photos/extract.py:47 (BUDGET_MODEL) |
openai/gpt-4.1-nano |
Photo description, budget tier (<30KB) | π¨ HARD SHUTDOWN 2026-10-23 β 21 DAYS. THE ONLY REAL DEADLINE ON THIS BOARD (rec 1, NEW). Verified at OpenAI's deprecations page today: gpt-4.1-nano / gpt-4.1-nano-2025-04-14 is deprecated with a shutdown date, vendor-recommended replacement gpt-5.6-luna. This is not a floor and not a notice-pending β it is a date, and calls fail after it. Unlike every other item on this list, doing nothing here breaks a working path. Proposed: google/gemini-3.1-flash-lite β already vita's budget-fast pin, proven at high volume, ~$0.10/$0.40 vs nano's $0.10/$0.40 (cost-neutral), and it adds no new vendor surface. In-family alternative is openai/gpt-6-luna ($0.10/$0.50, cheaper input than OpenAI's own suggested gpt-5.6-luna at $1/$6) β but it is still a WATCH row with no AA index, so prefer flash-lite. Also update MODEL_PRICING at :51-53 β unlike costs.py, that dict is a LIVE cost calculator for the models this script actually calls, not a historical ledger, so the new id needs a row or spend goes unrecorded. And PREMIUM_MODEL at :46 is google/gemini-3-flash-preview β a preview id two generations stale, no deadline, bump opportunistically to gemini-3.1-flash-lite or gemini-3.1-pro depending on whether photo quality needs the Pro tier. |
api/src/services/agenda_realtime.py:30 |
gpt-4o-mini-transcribe |
Agenda realtime transcription | πΆ DEPRECATED, removal 2027-02-27 (newly dated 2026-10-02). Verified at OpenAI today; replacements are gpt-live-transcribe or gpt-transcribe. ~5 months of runway, so not urgent β but now it has a date where before it had none. Bundle with rec 1 only if touching OpenAI ids anyway; otherwise revisit at the next refresh. Note this is a transcription model, so the replacement choice should be checked against whatever realtime/streaming mode this service actually uses β not swapped blind. |
scripts/feed/constitutions/extractor.py:34 + :72 |
claude-sonnet-5 + thinking={"type":"disabled"} |
Constitution amendment extraction | π HOLD β DO NOT BLIND-SWAP (rec 3). This site passes thinking={"type": "disabled"}, which adaptive-thinking Claudes reject with a 400 (fable-5/5.1 both 400 on it; opus-5-5 "thinking can no longer be disabled"), and sonnet-5-5 is adaptive. Deleting the param re-opens the exact bug the 09-21 pass fixed: adaptive thinking eats the max_tokens=400 budget and truncates the JSON mid-object. A swap here is a code change (raise max_tokens well above 400 and drop the param, then verify the JSON parses), not a one-line edit. Stays on sonnet-5 until that's done β sonnet-5 is Legacy but has no deprecation notice and a 2027-06-30 floor, so there is no clock. |
scripts/seo/geo.py:44 |
claude-sonnet-5 |
GEO/AI-search visibility panel | π HOLD β NOT a vendor API id (rec 3). These are DataForSEO model_name values; the file's own comment gates any change on GET /v3/ai_optimization/claude/llm_responses/models, and states that chat_gpt/gemini are deliberately held a few versions back because a bump breaks run-over-run comparability. The 09-21 bump here was forced by sonnet-4-5's retirement floor; sonnet-5's floor is 2027-06-30, so no forcing function exists and bumping would trade comparability for nothing. Revisit only if DataForSEO drops the model or a real floor appears. |
Audit snapshot: data/model-guidance/routing-audit-2026-10-02.json (25 pinned modes / 29 unpinned; 255 hardcoded refs, 75 stale, 45 unknown, 1 held). aloop_stale reads [] for the SECOND CONSECUTIVE RUN and is WRONG for the SAME unfixed reason β the 09-29 fix was a data edit to two structures and it did not happen. CURRENT_ROSTER (flows/model-guidance-review/snapshot_routing.py:29-73) still lacks claude-opus-5-5, claude-sonnet-5-5, gpt-6-sol, gpt-6-luna, deepseek-v4.1-flash β and now also gpt-6.1-sol and gemini-4-argon β while SUPERSEDED (:77-87) remains a full generation stale (claude-sonnet-4.6βclaude-sonnet-5 should be βclaude-sonnet-5-5; claude-opus-4.8/4.7βclaude-opus-5 should be βclaude-opus-5-5; gpt-5.5βgpt-5.6-sol should now be βgpt-6.1-sol). The consequence is unchanged and concrete: claude-sonnet-5 is in the roster and absent from the map, so both live sonnet-5 pins (config.json:39, :90) classify CLEAN while being this run's re-asserted recommendation. NEW BUG FOUND THIS RUN: .aloop/config.json:71 is DOUBLE-COUNTED β the sleep_build opus-4-7 pin is correctly bucketed held as an aloop mode and simultaneously stale as a hardcoded ref, because DELIBERATE_HOLDS (:100-103) matches on "aloop mode name or path substring" and .aloop/config.json is neither sleep_build nor scripts/cadence/daemon.py. So one deliberately-held pin inflates the stale count by one and will do so every run until the hold list carries the config path. AND THE 75 NEEDS TRIAGE, NOT ADOPTION β 12 of them are FALSE POSITIVES retired by reading one page. The audit tags 11 gpt-4o-mini refs plus 1 gpt-4.1-mini as "gpt-4 family (retired)"; the 09-23 and 09-29 reviews both deferred with "not verified at OpenAI this run β verify before sweeping". Verified at OpenAI's deprecations page today: neither gpt-4o-mini nor gpt-4.1-mini is deprecated, retired or dated. Do not sweep them. That caveat is CLOSED. What the same fetch did find is the run's only hard deadline β gpt-4.1-nano shuts down 2026-10-23 (live at photos/extract.py:47) and gpt-4o-mini-transcribe is removed 2027-02-27 (live at agenda_realtime.py:30). A zero is not evidence until the filter is proven able to return something β and a 75 is not evidence until you read what is in it. Remaining genuine clusters, unchanged in substance: 33 sonnet-4.6 refs (voice-critical, deliberately held pending a drift check β now held twice over, since sonnet-5-5 exists and the pilot has not been re-pointed), 7 minimax-m2.7 (the m3 bake-off that has never been run), ~18 gemini-2.5/3.0-era refs in screenjournal/, photos/extract.py, drafts/, documents/scan_filer.py, and the stale research-proposal panel above. Prior note (2026-09-29), kept for the reasoning trail: aloop_stale reads [] again β and this time it is WRONG, for a NEW reason. The 09-21 structural fix holds (classification is roster-driven, not regex-driven), but the roster itself has fallen behind by maintenance lag: known_roster is missing claude-opus-5-5, claude-sonnet-5-5, gpt-6-sol, gpt-6-luna, deepseek-v4.1-flash β every model rostered on 09-23 and 09-29 β and superseded_map is a full generation stale (claude-sonnet-4.6βclaude-sonnet-5 should now be βclaude-sonnet-5-5; claude-opus-4.8/4.7βclaude-opus-5 should be βclaude-opus-5-5; gpt-5.5βgpt-5.6-sol should be βgpt-6-sol). The consequence is concrete: claude-sonnet-5 is in known_roster and is not a key in superseded_map, so both live sonnet-5 pilot pins (config.json:39, :90) classify as CLEAN when they are in fact stale β the audit could not have surfaced this run's headline recommendation. A zero is not evidence until the filter is proven able to return something. Fix is rec 5: it is a data edit to two structures, not a rewrite. Note this is a different bug from the 09-21 one β that was structural (a frozen hand-maintained regex list); this is a maintenance process that has no trigger tying roster updates to roster additions. Two long-carried items CLOSED this run (both verified in code, not assumed): the six downstream council flows are fully synced to opus-5-5/gpt-6-sol, and the cost-tracking gap β carried for 13 reviews β is closed: scripts/inference/costs.py:27-38 now holds opus-5-5/opus-5/fable-5-1/sonnet-5 with correct per-model cache multipliers and an explicit "cache_read is NOT a fixed fraction of input β never derive it" comment. Only claude-sonnet-5-5 needs adding (rec 4; identical rates to sonnet-5). claude-haiku-4-5 is no longer a vita concern despite its 10-15 floor: zero live code pins remain (the three were routed to sonnet-5/flash-lite on 09-21) β the only hits are data/tokens/*.json historical usage records, which are catalog by design. Prior note (2026-09-23): The fixed audit is now earning its keep β and the 09-23 run is the first time it produced a genuinely new, actionable finding rather than a re-list: the 88 stale refs are no longer "all documented residuals." The real cluster is six downstream council flows still on opus-4-7/opus-4-8 (see the routing row above), which the old frozen-roster gauge could never have seen because it listed opus-4-7 as CURRENT. Other genuine clusters worth a decision: 26 sonnet-4.6 refs (16 claude-sonnet-4.6 + 10 claude-sonnet-4-6) β these are the voice-critical paths deliberately held while the sonnet-5 pilot runs, NOT neglect, but with Sonnet 5.5 signalled "within weeks" the hold is now the right call twice over; 11 gpt-4o-mini refs in gumball backend services + scripts/photos/extract.py (classified retired-family by the audit β not verified at OpenAI this run, so verify before sweeping); 8 gpt-5.5 refs (the council seats above + seo/geo.py:42); 7 minimax-m2.7 (the never-run m3 bake-off targets); and ~18 gemini-2.5/3.0-era refs in screenjournal/, photos/extract.py, drafts/, documents/scan_filer.py. aloop_stale is legitimately [] this run β all 25 pinned modes check out against the live roster, and the one opus-4-7 pin (sleep_build) is correctly bucketed held, not hidden. Prior note (2026-09-21): THE FROZEN-ROSTER BLIND SPOT IS FIXED (11th flag was the last one). snapshot_routing.py no longer classifies staleness from a hand-maintained regex list that could not see anything past opus-4-7. It now canonicalizes each id (vendor prefix, date suffix, -preview, 4-6β4.6) and checks it against CURRENT_ROSTER + a SUPERSEDED map, with four buckets: stale (superseded/retired β actionable), off-roster (unrecognized, mostly changelog prose), held (the DELIBERATE_HOLDS allowlist: sleep_build / cadence/daemon.py / generate_html_v2.py on opus-4-7), and catalog (CATALOG_PATHS β price tables, usage parsers, containment-test pins, which name old models by design). Fixing the roster exposed a second bug in the same file: GREP_PATTERN truncated ids at the first word (gemini-3.1-flash-lite β gemini-3.1-flash), which the old prefix-regexes tolerated and roster comparison did not β 244 refs found post-fix vs 227. Post-fix run: 0 aloop stale Β· 88 hardcoded stale Β· 26 off-roster Β· 4 held. The deep-council:28 / gpt-5.6-sol false positives are gone β sol is on the roster now, so it reads ok. The 09-21 genuine live stale pins (geo.py:41 sonnet-4-5, setup_models.py:5 opus-4-7) and all three haiku-4-5 pins were applied the same day; see the Change log row.
Sonnet 4.6 swap decisions (2026-04-23)
Sonnet 4.6 was being used heavily in both agentic and non-agentic paths. Audited every use site and swapped the non-voice-critical ones to cheaper Chinese models of equal-or-better benchmark performance on those specific workloads. Voice-critical paths stayed on Sonnet.
| Site | From | To | Rationale |
|---|---|---|---|
flows/code-review/synthesize |
sonnet-4-6 | z-ai/glm-5.1 | Code reasoning + structured output; GLM-5.1 beats Sonnet on SWE-bench Pro (58.4 vs ~55), ~1/3 cost |
flows/council-plan/council |
sonnet-4-6 | z-ai/glm-5.1 | Architectural review of implementation proposals; pure code-reasoning |
flows/docs-audit/review |
sonnet-4-6 | moonshotai/kimi-k2.6 | 1M context lets it read more docs at once |
flows/council-review/synthesize |
sonnet-4-6 | moonshotai/kimi-k2.6 | Long-horizon synthesis across multiple model outputs |
flows/podcast-deep/quality_gate.py |
sonnet-4-6 | minimax/minimax-m2.7 | Structured yes/no judging with rubric; M2.7 adequate, cheap |
flows/vdb/generate_newsletter.py |
sonnet-4-6 | moonshotai/kimi-k2.6 | Long-form data-driven newsletter |
scripts/memory/consolidator.py |
sonnet-4-6 | moonshotai/kimi-k2.6 | Weeklyβmonthly rollup, long context |
scripts/feed/narrative.py |
sonnet-4-6 | z-ai/glm-5.1 | Structured editor pass over timeline entries |
scripts/feed/privacy.py |
sonnet-4-6 | z-ai/glm-5.1 | Structured editor pass for privacy scrubbing |
api/src/services/improve_cycle.py |
sonnet-4-6 | z-ai/glm-5.1 | Improvement proposal generation = code/system reasoning |
gumball/flows/field-guide/segments/signal-check/FLOW.yaml |
sonnet-4-6 | moonshotai/kimi-k2.6 | PILOT β 1 segment only; expand to other segments if quality judgment is positive |
Kept on Sonnet 4.6 (voice-critical):
- flows/podcast-deep/scratchpad β creative brainstorm for the podcast host voice
- flows/podcast-quick/transcript β actual transcript output
- scripts/coaching/generator.py β coaching messages to bio-zack
- scripts/podcast/generate_transcript.py β podcast scripting output
- 6 of 7 gumball/flows/field-guide/segments/* + field-guide-review.flow.yaml β hold until signal-check pilot judged
Estimated monthly savings: ~$20-40 at current volumes. Real value is concentrating Sonnet "quality budget" on voice surfaces where drift is hardest to measure and easiest to miss.
Observability: token usage by model is already logged to data/tokens/YYYY-MM.json and data/inference/usage.jsonl. Watch for quality regressions via (a) bio-zack reports, (b) flows/podcast-deep/quality_gate.py failure rate (now self-judging), (c) /improve proposal quality trend.
Residual refs (intentional, do not touch)
scripts/inference/costs.py,cost_report.py,usage.pyβ keep historical model IDs (claude-opus-4-5-20251101,gpt-5.4, etc.) for retrospective cost computation on token data. New IDs added alongside, not replacing.scripts/tokens/codex_extractor.pyβ parses old codex logs that referencegpt-5.4; intentional.scripts/telegram_bot.pyβminimax-m2.5-faston SambaNova (may not have M2.7-fast hosting yet; bump when available).flows/containment-*/FLOW.yamlβkimi-k2-0905,kimi-k2.5pinned for reproducibility in security boundary tests.
By-task recommendations
Use this as a decision tree when writing a new script, a new stepwise step, or a new .aloop mode.
Real-time chat (telegram, voice, silicon-zack sessions)
β claude-sonnet-4-6 default; escalate to claude-opus-4-8 if the turn involves coding or a multi-step plan. Do not route to OpenAI from the chat path β voice fidelity is tuned to Claude. Sonnet 5 pilot (2026-07-05): do NOT swap chat/voice/composer paths to sonnet-5 yet β pilot it first on a non-voice analytical path (media_deep_dive), read quality, and account for the ~1.0-1.35Γ tokenizer inflation in cost before touching any voice-critical surface.
Proactive composers (scripts/proactive/builders.py, commitment_followup, morning_kickoff)
β claude-sonnet-4-6. Pricing matters here β these fire dozens of times/day. Avoid Opus unless a specific composer has shown quality issues.
Executive loop / sleep reflect / identity work
β claude-sonnet-4-6 for reflection passes (cheaper, runs often), claude-opus-4-8 for the sleep build step and any autonomous-edit agent (VITA #1 issue β persistence quality matters most here).
Coding agents (sleep build, stepwise agent executors, /improve develop)
β claude-opus-5-5 (2026-09-22) is the new default β $4/$20 (20% under opus-5's $5/$25), AA Intelligence Index #1 at 58 on v4.3.2 with a 5-point lead over Fable 5.1 and Astra, leading 6 of 10 components, ZDR-clean, 1M ctx model-card-confirmed. Wire the opus-4-8 cyber fallback (5.5 redirects most cyber tasks there). Two things to carry when you adopt it: (a) it is ~1.6Γ more verbose, so AA's cost-per-task is FLAT vs opus-5 ($5.98 vs $5.86) β adopt it for the quality, not for a 20% saving; budget agentic paths as if the price were unchanged and take the win on cache-heavy/short-output paths instead; (b) thinking cannot be disabled and medium is the default effort β any caller that relied on opus-5's default or on thinking:{type:disabled} (e.g. the small-max_tokens JSON pattern used at feed/constitutions/extractor.py) will behave differently; set effort explicitly. It supersedes claude-opus-5, which is retained as the last un-watermarked frontier Claude (use it for anything published AS bio-zack where the EU-AI-Act watermark matters) and at any not-yet-swapped pin, and claude-opus-4-8, which stays as the cyber/availability fallback. gpt-6-sol (2026-09-22) is the new cost-per-task challenger in this lane β $2/$10, AA 48 vs Opus 5.5's 58, but $0.13β$1.06/task against $5.98. For bounded coding where 10 index points are affordable, it is now a stronger pilot arm than grok-4.7 ($3.74/task at AA 46) β fold it into the still-unrun bounded-coding pilot rather than opening a new one. Use kimi-k2.6 when cost matters and the task is bounded. deepseek-v4-pro is a strong batch-mode alternative β slow output (34 TPS) so async only. glm-5.2 is now the strongest open-weight coding option (SWE-bench Pro 62.1, top open-weights, MIT) for non-sensitive code β usable 1M ctx, high/xhigh thinking, but async-friendly only (TTFT ~10.8s); glm-5.1 stays as the ~30%-cheaper fallback for high-volume non-coding structured passes (feed editors). For the hardest, highest-value builds where correctness dominates cost, claude-fable-5-1 now leads (supersedes fable-5 09-01; AA #1 at 65.7 v4.1.1, Terminal-Bench-Science 52.6, GDPval 1853 β same $10/$50 but 75% cheaper cache reads + ~25% cheaper effective) β but at 2Γ opus-5's output cost, presumed non-ZDR, now watermarked, and with mandatory refusal/fallback handling, it's a deliberate per-task escalation, not a default. Pilot candidate for sleep_build only if a build ever needs more than opus-5 (the opus-4-7 pins stay a deliberate hold; opus-5 is the eventual target there). kimi-k2.7-code (2026-07-05): coding-specialized k2.6 successor, ~30% fewer thinking tokens and cheaper output ($0.74/$3.50) β but 256K ctx and vendor-only benchmarks. RETAIN k2.6 on all its current sites (they're 1M-context synthesis roles k2.7-code can't serve); rostered for future bounded coding agents where β€256K suffices β no current vita site is a clean swap target. grok-4.5 (2026-07-09): cheaper-frontier coding/agentic model whose value prop is cost-per-task β $2/$6 with ~2Γ token efficiency (4.2Γ fewer output tokens than Opus 4.8 on SWE-bench Pro), AA Coding Agent Index 76, and it beats Opus 4.8 on Terminal Bench 2.1 + SWE Marathon while trailing on SWE-bench Pro. Pilot only, NOT auto-swapped: run it as a challenger on ONE bounded coding path (candidate: flows/code-review/synthesize, currently deepseek-v4-pro) and measure effective cost-per-task WITH the ~2Γ token efficiency applied β per-token rates alone understate it. Keep it OFF any >500K-context role (500K ctx regression) and OFF council (no Omniscience number). Proprietary (not open-weight), so it competes with opus-4-8/gpt-5.5, not glm/kimi. grok-4.6 (2026-08-12) SUPERSEDES grok-4.5 as the pilot arm: same $2/$6, +5 AA Intel Index (61), and an even better efficiency profile (~ΒΌ the input tokens of Opus 5 per AA) β so if the grok-4.5 bounded-coding pilot on flows/code-review/synthesize hasn't run yet, run grok-4.6 instead (update scripts/experiments/model_dialogue.py:29 GROK = "x-ai/grok-4.5" β grok-4.6 and the scripts/council/core.py:25 registry entry). Same guardrails: measure effective cost-per-task (per-token rates understate it), keep it OFF >200K-context work (rebills at $4/$12) and OFF council. Reality-check on coding specifically: AA is explicit that grok-4.6's strength is agentic/knowledge-work + cost, and its coding rows (DeepSWE 65.9) TRAIL GPT-5.6 Sol Max β so this is a cost-per-task pilot, not a quality-leader swap; do not read "AA Index 61 = Sol" as coding parity. gpt-5.6-luna (2026-07-10): add as a THIRD arm to that same bounded pilot β $1/$6, Terminal-Bench 2.1 84.7 (above grok-4.5's 83.3 and Opus 4.8), Coding Agent Index 74.6, AA-verified. One pilot, three challengers (v4-pro incumbent vs grok-4.5 vs luna), one cost-per-task readout β beats running two separate pilots. Luna's hard constraint: MRCR 41.3% β the task must fit comfortably under ~100K context. kimi-k3 (2026-07-18): WATCH β no coding routing yet. Frontier-tier open-weight (AA v4.1 57) whose only independently-verified coding signal is Arena.ai frontend-code #1, ahead of Fable-5 β so the natural future pilot is a bounded FRONTEND-generation path (the HTML-report implementer in scripts/reports/generate_html_v2.py, or /frontend-slides), NOT the SWE-style bounded-coding pilot lane (where all its numbers are vendor-harness). Three gates before any pilot: (1) weights + license land (~Jul-27), (2) OpenRouter capacity stabilizes (frequent 429s today), (3) at least one harness-comparable independent coding number. At $3/$15 it competes with sonnet/terra/grok-4.5 β not with the glm/kimi cheap band, and it does NOT touch k2.6's slots.
Research, synthesis, summarization (research_*, synthesis, deep_analysis modes)
β minimax-m2.7 is the incumbent (Intel Index 50, $0.30/$1.20). deepseek-v4-flash is the active challenger ($0.14/$0.28, Intel Index 47, 1M context) β bake-off in progress on flows/podcast-deep/quality_gate.py and scripts/media/synthesis.py. If quality holds, fan out to the full M2.7 tier. minimax-m3 (2026-07-05): the direct m2.7 successor at the SAME $0.30/$1.20 but with 1M ctx + native multimodal β add it as a third bake-off arm and pilot it on the equal-price m2.7 hardcoded sites (scripts/speculative/{engine,storage,budget}.py); it's a capability upgrade at zero price delta over m2.7 (though still pricier than v4-flash). Verify the OpenRouter rate is permanent (not promo) first. gpt-5.6-luna is NOT a budget-band candidate (2026-07-10): at $1/$6 it's 3β5Γ m2.7's sticker and ~20Γ v4-flash's output rate β its niche is cheap frontier-agentic coding (see Coding agents), not high-volume synthesis; and its MRCR 41.3% disqualifies the long-context synthesis roles this band leans on. kimi-k3 likewise is NOT a candidate here (2026-07-18): $3/$15 is 10Γ m2.7's sticker β its lane is frontier-value, and k2.6 keeps ALL the 1M-ctx strong-cheap synthesis slots.
High-volume triage (media scoring, daily rollup, trend analysis)
β gemini-3.1-flash-lite stays. Nothing touches its $/M.
Vision / OCR / screenshot analysis (emails with receipts, Libre CGM screenshots, health screenshots)
β claude-opus-4-8 for text-in-image OR gemini-3.1-pro for PDFs + diagrams. Flash-lite is not good enough here. qwen3.8-max (2026-08-05): WATCH β the most interesting future challenger in THIS lane. #2 Vision Arena + native text/image/VIDEO input at $2/$6 (vs gemini-3.1-pro's $2/$12) is a genuinely differentiated cheap-multimodal profile β but Arena rank β an OCR/PDF/receipt benchmark, and vita's vision paths touch health/CGM/labs (China data-jurisdiction risk applies). Do NOT route any vision work here until (a) an AA/independent OCR-or-document-VQA number exists and (b) weights land so sensitive paths can self-host. Pilot candidate only, gated β see Open Questions.
Council / delphi / tribunal (high-stakes decisions)
β Four-seat fan-out. PROPOSED 2026-09-23 (rec 2 + 4): Claude seat + synthesis β claude-opus-5-5, OpenAI seat β gpt-6-sol. The Claude swap is opus-5 β opus-5-5: 20% cheaper per token, AA #1 by 5 points, ZDR-clean, and it leads AA-Omniscience β the honesty component this section has used as the seat gate since 2026-04. The OpenAI swap replaces the obsolete 07-10 "gpt-5.5 β gpt-5.6-sol" proposal: gpt-6-sol shipped 09-22 at $2/$10 vs gpt-5.5's $5/$30, so the seat gets two generations newer and ~65% cheaper in one edit. The METR reward-hacking flag was raised against gpt-5.6-sol and has NOT been re-checked for gpt-6-sol β confirm at metr.org before or shortly after. Current live seats: claude-opus-5, gpt-5.5, gemini-3.1-pro, grok-4.20; synthesize with the Claude seat. Reserve gpt-5.5-pro for rare, bio-zack-tagged critical calls. β grok-4.20's seat needs re-examination for the first time (2026-09-23, bio-zack's call β see Open Questions): its $20/$60 sticker is justified entirely by an AA-Omniscience honesty lead, and Opus 5.5 now leads that component. That does not automatically retire it β AA's own framing is that Anthropic's Omniscience standing is calibration-driven (low hallucination) while xAI's is accuracy-driven (highest % correct, high hallucination), so they are genuinely different angles, and a council loses value if two seats become one vendor. But the seat has never been challenged before and it is the most expensive line in vita's routing. Pull the isolated Opus-5.5-vs-grok-4.20 Omniscience numbers on the same scale before deciding. PROPOSED 2026-07-25 (not yet applied): bump the Claude deliberation seat AND the synthesis model opus-4-8 β claude-opus-5 β same $5/$25, #1 on AA Intelligence Index (61 v4.1, above Fable-5's 60), joint-#1 Coding Agent Index, record GDPval/AA-Briefcase. This SUPERSEDES the 06-17 Fable-5-synthesis pilot proposal: opus-5 gets you Fable-5-or-better synthesis quality at HALF the cost, is ZDR-clean (no 30-day-retention block), and carries no export-control availability risk β so it's a cleaner synthesis upgrade than Fable-5 on every axis except raw verbose deliberation. Wire the opus-4-8 cyber-refusal fallback (health/doctor-adjacent councils were already OFF Fable-5 for classifier risk; opus-5's classifiers fire ~85% less but keep the fallback). Sites: flows/council/setup_models.py, scripts/council/core.py (audit-flagged: core.py:22 still pins opus-4-7, not opus-4-8 as this doc claimed β reconcile to opus-5), + the 5 downstream flows (simple-council, council-review, deep-council, research/codebase-research, vision-strategy). One change at a time: land the Claudeβopus-5 seat swap independently of the still-open gpt-5.5βgpt-5.6-sol OpenAI-seat proposal below. PROPOSED 2026-07-10 (not yet applied): bump the OpenAI seat gpt-5.5 β gpt-5.6-sol β same $5/$30, successor generation, stronger agentic/reasoning numbers (v4.1 59 vs Fable-5's 60), GDPval β Fable-5. Sites: flows/council/setup_models.py, scripts/council/core.py, + the 5 downstream flows (simple-council, council-review, deep-council, research/codebase-research, vision-strategy). Two reasons this is a proposal and not an auto-swap: (a) the METR reward-hacking report is unconfirmed at primary, (b) Omniscience shows hallucination rate UP vs 5.5 β acceptable for the reasoning seat since grok-4.20 holds the honesty seat, but worth bio-zack's eyes. Keep gpt-5.5-pro as-is (no 5.6-pro exists). Grok-4.3 shadow-seat pilot concluded β see open questions for retirement decision status. grok-4.5 evaluated for a seat (2026-07-09) β DECLINED: it publishes no AA-Omniscience component, and grok-4.20's seat rationale is entirely its 78% Omniscience honesty lead β grok-4.5 adds cheaper capability, not a distinct honesty perspective, and would be redundant with the xAI slot. Keep the 4 seats; re-evaluate only if AA publishes an Omniscience number for it. grok-4.6 evaluated for a seat (2026-08-13) β DECLINED, same test: still no published AA-Omniscience component, so it adds intelligence/cost but not a distinct honesty angle β grok-4.20 keeps the honesty seat. Council roster UNCHANGED this run. gpt-6-astra evaluated for a seat (2026-09-04) β DECLINED, same test: no AA-Omniscience component, $10/$50 (2Γ the incumbent OpenAI seat), and a 51%-at-max-effort hallucination rate (AA) β it adds cost, not a distinct honesty angle; grok-4.20 keeps the seat. Council-code reconciliation (audit 09-04): scripts/council/core.py:22/33 now correctly pin claude-opus-5 and :23 gpt-5.6-sol β the 07-25 opus-4-7 drift is RESOLVED in core.py; the remaining open item is confirming flows/council/setup_models.py matches (the routing table above still shows opus-4-8/gpt-5.5 there β so the standing opus-4-8βopus-5 Claude-seat + synthesis swap may still be pending on THAT path, and gpt-5.5βgpt-5.6-sol likewise). kimi-k3 evaluated for a seat (2026-07-18) β DECLINED: same test β no Omniscience component published, upstream capacity-limited (frequent 429s β unfit for a reliability-sensitive fan-out), and its distinct angle (open weights, price) is not a council criterion. Re-evaluate only after weights land AND an honesty number exists. Fable-5 synthesis pilot (proposed 2026-06-17): the synthesis step is a single, low-volume, highest-stakes call β the cleanest place to spend Fable-5's 2Γ cost. Recommend piloting claude-fable-5 as the synthesis model only (keep the 4 deliberation seats), once council code carries refusalβopus-4-8 fallback. Health/doctor-adjacent councils stay OFF Fable-5 β bio classifier false-positive risk.
Embeddings / classifiers / small utility models
β Out of scope for this doc β tracked in scripts/vdb/ and scripts/inference/. Revisit when a frontier-embedding model ships.
Provider access matrix
| Provider | Env var | SDK | OpenRouter proxy? | Notes |
|---|---|---|---|---|
| Anthropic | ANTHROPIC_API_KEY |
anthropic python |
yes | Preferred direct; Claude daemon uses this. claude-opus-5-5 (GA 09-22, $4/$20, cache read $0.20 at a 0.05Γ multiplier, batch $2/$10, fast $8/$40, 1M ctx / 128K out, effort default medium, ZDR available) SUPERSEDES claude-opus-5 β live on Claude API + Claude Platform on AWS + Bedrock (anthropic.claude-opus-5-5) + Google Cloud + Azure/Foundry + OpenRouter day-one (anthropic/claude-opus-5.5). It is watermarked (EU AI Act) and cannot disable thinking; it redirects most cyber tasks to opus-4-8. Note the anti-distillation safeguard on API accounts created after 2026-08-31 β vita's predates it, but confirm. claude-opus-5 is now Anthropic-"Legacy" (no deprecation notice; retirement floor 2027-07-24) and is retained as the last un-watermarked frontier Claude. Prior: claude-opus-5 ($5/$25, ZDR-clean, 2Γ fast-mode); OpenRouter availability unverified at the 07-25 run. opus-4-8 retained as the cyber-refusal fallback for opus-5/Fable-5 callers. claude-fable-5-1 (GA 09-01, $10/$50, cache-read $0.25, batch $5/$25, 1M ctx/128K out) SUPERSEDES claude-fable-5 on Anthropic API + Bedrock + Google Cloud + Azure/Foundry; AA #1 (65.7). NEW: can discover vulns (not exploits) β GA authorized-security path; and embeds an invisible EU-AI-Act text watermark (detection is Anthropic-API-only). ZDR posture unconfirmed β verify before sensitive routing. claude-mythos-5-1 = trusted-access US-orgs-only, not adoptable. |
| OpenAI | OPENAI_API_KEY |
openai |
yes | Direct for GPT-5.5-pro (response-API feature parity). gpt-5.6 family (gpt-5.6-sol/-terra/-luna; bare gpt-5.6βSol) live day-one direct + OpenRouter; programmatic tool calling is Responses-API-only β direct SDK required for that feature. NEW cache billing on 5.6+: writes 1.25Γ input, reads β90%, 30-min min life. gpt-6-astra (GA 09-03, $10/$50, cached $1, cache-write $12.50, 1.05M ctx/128K out, cutoff 2026-04-30) is OpenAI-API-ONLY at launch (NOT on OpenRouter) β direct SDK required; >272K-token prompts rebill the whole request to ~$20/$75. WATCH, not routed. Sol NOT deprecated (promo $4/$0.40cached/$20 through 2026-11-21). |
GEMINI_API_KEY |
google-genai |
yes | Flash-lite (gemini-3.1-flash-lite) used heavily for budget-fast triage β retained. gemini-3.1-pro = frontier-multimodal. google/gemini-3.7-flash (GA 08-13, 1M ctx, multimodal, OpenRouter day-one) is a new WATCH mid-tier β intro $0.75/$3.75 through year-end β $1.50/$7.50 in 2027, but OpenRouter surfaced $0.375/$1.875 (verify at model card before volume). Not a flash-lite replacement. Direct preferred for cost tracking. |
|
| xAI / SpaceXAI | XAI_API_KEY |
OpenAI-compatible | yes | Provider rebranded xAI β SpaceXAI (07-08 grok-4.5 launch) β confirm base-URL/key path unchanged. OpenRouter fine β low volume (council only). x-ai/grok-4.6 (GA 08-12, $2/$6, 500K ctx, Feb-1-2026 cutoff) SUPERSEDES x-ai/grok-4.5 at the same base price. Tiered pricing: prompts >200K tokens rebill the whole request at $4/$12 (grok-4.6 + grok-4.3 both) β verify the higher band before routing long-context jobs; grok-4.6 OpenRouter weighted-avg input ~$0.74 (~90% cache-hit). |
| MiniMax | β | β | via OpenRouter only | minimax/minimax-m2.7 (incumbent); minimax/minimax-m3 (successor, same price, 1M ctx + multimodal β pilot) |
| Z.ai | β | β | via OpenRouter only | z-ai/glm-5.2 (current strong-cheap-code pin, $1.40/$4.40); z-ai/glm-5.1 retained as cheaper fallback. z-ai/glm-5.3 (GA 08-14) SUPERSEDES 5.2 but is WATCH-only β per-token price UNPUBLISHED (GLM Coding Plan subs only), NOT on OpenRouter, open weights + license ~2wk out (do not assume MIT carries). Cyber lane (CyberGym 84.5) is the vita-relevant watch. China data-jurisdiction risk on API path β non-sensitive code only, or self-host the weights once they land |
| Moonshot | β | β | via OpenRouter only | moonshotai/kimi-k2.6 (1M ctx, general β retained); moonshotai/kimi-k2.7-code (256K, coding-specialist β pilot); moonshotai/kimi-k3 ($3/$15, cache-hit $0.30, 1M ctx, text+image β WATCH: single-provider, upstream capacity-limited w/ frequent 429s, weights promised 07-27) |
| DeepSeek | DEEPSEEK_API_KEY |
OpenAI-compatible | yes | V4-Pro + V4-Flash live since 2026-04-24. OpenRouter ok; direct preferred for V4-Flash if volume warrants (saves 5% markup) |
| Alibaba / Qwen | β | β | via OpenRouter (single provider) or DashScope | qwen/qwen3.8-max ($2/$6, cache-hit implicit $0.25 / explicit-read $0.17, 1M ctx, text+image+video β WATCH): OpenRouter live day-one but single-provider (capacity risk like K3). DashScope = Alibaba Cloud Model Studio direct (China data-jurisdiction / state-law risk β non-sensitive only). Open weights ~week of Aug-10 (license TBD) β self-host path opens then. Verify OpenRouter $2/$6 vs the Qubrid $2.87/$7.12-list outlier before any volume routing. |
OpenRouter key lives at OPENROUTER_API_KEY. All non-direct providers route through scripts/inference/factory.py.
Change log
Each refresh appends a row. Oldest-first β tail is freshest.
| Date | Change | Triggered by |
|---|---|---|
| 2026-04-23 | Initial doc. Opus 4.7 adopted for frontier-agent. Kimi K2.6 added to strong-cheap tier. M2.7 flagged as bump from M2.5. Deprecated Sonnet-4 references in .aloop/config.json. |
Manual build by bio-zack request |
| 2026-04-23 | Sonnet 4.6 β cheaper-Chinese swap pass (bio-zack directive): 5 stepwise flows + 5 scripts + 1 API service + 1 gumball pilot segment moved off Sonnet 4.6 onto GLM-5.1 (code-reasoning), Kimi K2.6 (long-horizon synthesis + 1M context), or MiniMax M2.7 (cheap structured judging). Voice-critical paths (coaching/generator, podcast transcripts, podcast-deep scratchpad, 6/7 gumball segments + field-guide-review) kept on Sonnet. See "Sonnet 4.6 swap decisions" section above for per-site rationale. | bio-zack "just do it all and ill watch" |
| 2026-04-23 | Applied all 6 recommendations: (1) .aloop research/background/deep_analysis/synthesis/speculative/research bumped M2.5βM2.7; (2) .aloop epd47_generate + media_deep_dive bumped sonnet-4βsonnet-4-6; (3) added .aloop mode sleep_build pinned to opus-4-7; (4) council roster bumped to opus-4-7/gpt-5.5/gemini-3.1-pro/grok-4.20 in flows/council/setup_models.py + scripts/council/core.py + 5 downstream flows (simple-council, council-review, deep-council, research-v2, vision-strategy); (5) swept 19 script + flow files onto current IDs (memory/consolidator, coaching/generator, cadence/daemon, feed/narrative, feed/privacy, podcast/generate_transcript, improve_cycle, vdb/generate_newsletter, speculative/{budget,storage,engine}, sentinel/poll_agent, media/{deep_analysis,synthesis}, podcast-panel-reactive); (6) DeepSeek V4 tracked in Open Questions. Also: tightened snapshot regex to capture full model strings (fewer false positives); added cost-table entries for opus-4-7. Aloop stale: 11β0. Hardcoded-stale true positives eliminated; 28 residual flags are intentional historical cost-tracking + containment-test pins. |
bio-zack "implement all suggested feedback" |
| 2026-04-23 | Automated audit: confirmed 11 stale aloop modes (8Γ m2.5βm2.7, 2Γ sonnet-4β4.6, 1Γ sleep unpin). 7 stale hardcoded sonnet-4 refs in scripts. Council seats (opus-4.6+gpt-5.4) and 5 Kimi K2 flow refs flagged for bump. 30+ audit false positives identified (pattern matcher too greedy on model-family substrings). No new models (research skipped). Full report: data/reports/2026-04-23-model-guidance-review.md. |
flows/model-guidance-review automated run |
| 2026-04-30 | DeepSeek V4 Pro + Flash adoption pass. Both shipped 2026-04-24 (MIT-licensed, 1M context, OpenRouter-available). V4-Pro added to Strong-cheap-reason tier (Intel Index 52, SWE-bench Verified 80.6%, Codeforces 3206 β frontier-tier reasoning at ~1/6th Opus 4.7 cost). V4-Flash added as Budget-MoE challenger to M2.7 (Intel Index 47, $0.14/$0.28 β exactly matches doc's prior projection). Domain scoring table extended with both. 75% promo on V4-Pro through 2026-05-05 β pilot recommendation: 2-flow V4-Flash bake-off vs M2.7 (podcast-deep/quality_gate.py, media/synthesis.py); add V4-Pro as 5th council seat through promo expiry; pilot V4-Pro on flows/code-review/synthesize (replace GLM-5.1). Audit: 0 stale aloop modes, 28 hardcoded refs all intentional residuals. Full report: data/reports/2026-04-30-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected |
| 2026-04-30 | Grok 4.3 review (later same day). xAI Grok 4.3 went GA today (beta since Apr 17), priced $1.25/$2.50 per MTok β 16x/24x cheaper than grok-4.20 β with Intel Index 53 (+5) and #1 output speed (206.9 TPS). xAI did not publish a model card, so SWE-bench / GPQA / AA-Omniscience are unavailable β and grok-4.20's roster slot is justified entirely by its 78% Omniscience honesty lead. Decision: add grok-4.3 to the roster as a "Reasoning-cheap (xAI)" tier with SHADOW-PILOT status; recommend a 5th council seat through ~2026-05-14 to gather side-by-side honesty signal. Hold on grok-4.20 retirement until AA publishes the Omniscience component for 4.3. Domain scoring table extended with conservative scores + explicit "?" markers where benchmarks are missing. Audit re-run found 0 stale aloop modes; 28 hardcoded refs remain (all intentional cost-tracking + containment-test residuals). Full report: data/reports/2026-04-30-model-guidance-review-grok-4-3.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (x-ai/grok-4.3) |
| 2026-06-18 | GLM-5.2 lands β Strong-cheap-code tier bumped glm-5.1βglm-5.2. Z.ai/Zhipu GLM-5.2 (GA 2026-06-16, 753B open weights, MIT, $1.40/$4.40, 1M usable ctx via IndexShare, high/xhigh thinking) is the top open-weights coding model: SWE-bench Pro 62.1 (> GPT-5.5 58.6 > glm-5.1 58.4), FrontierSWE 74.4% / MCP-Atlas 77.0 near-tie Opus 4.8. Updated roster + domain scoring (coding 62.1% Pro; long-horizon 7β8; long-ctx 7β9; tool-use 8β9; cost-eff 10β9; streaming 9β7 for ~10.8s TTFT thinking latency). Routing proposals (NOT auto-applied): (1) bump 2 code-reasoning sites glm-5.1βglm-5.2 (flows/council-plan/FLOW.yaml:19, api/src/services/improve_cycle.py:68); (2) HOLD 2 high-volume feed-editor sites on glm-5.1 (scripts/feed/narrative.py:144, scripts/feed/privacy.py:170) β coding gains don't apply, 5.2 is ~30% pricier; (3) re-surface the still-unapplied 2026-06-17 opus-4-7β4-8 bump (sleep_build aloop + scripts/cadence/daemon.py:258; also scripts/reports/generate_html_v2.py:54-55); (4) close cost-tracking gap β add opus-4-8/fable-5/glm-5.2 rows to scripts/inference/costs.py + cost_report.py. Audit: 0 stale aloop modes; 26 hardcoded "stale" hits all documented intentional residuals (cost-history, containment-test pins, SambaNova-gated m2.5-fast TODO, reflect.md placeholder). Benchmarks from Z.ai launch coverage β NOT yet on artificialanalysis.ai; AA Intel Index unpublished; re-confirm next refresh. Full report: data/reports/2026-06-18-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (glm-5.2) |
| 2026-06-17 | Claude Fable 5 lands β new Frontier-max tier. Added claude-fable-5 (Anthropic, GA 2026-06-09, $10/$50, 1M ctx, SWE-bench Pro 80.3 β new top reasoner) to roster + domain scoring table as a tier ABOVE opus-4-8; reserved for hardest builds + council synthesis, NOT a default (2Γ Opus output cost). Flagged Fable-5's integration requirements (adaptive-thinking-always-on, refusal/fallback opt-in, 30-day-retention/no-ZDR) and its sibling claude-mythos-5 (Glasswing-only, not adoptable). Routing proposals (NOT auto-applied): (1) pilot Fable-5 as council synthesis model; (2) bump live opus-4-7βopus-4-8 on aloop sleep_build + scripts/cadence/daemon.py:258; (3) add opus-4-8 + fable-5 pricing rows to scripts/inference/costs.py + cost_report.py. Audit: 0 stale aloop modes; hardcoded "stale" hits are documented intentional residuals (cost-history, containment pins, SambaNova-gated m2.5-fast TODO, reflect.md placeholder). Full report: data/reports/2026-06-17-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (claude-fable-5) |
| 2026-06-23 | gpt-5.5-cyber evaluated β NOT adoptable; no roster change. OpenAI's cyber-specialist (full version ~2026-06-22; CyberGym 85.6 vs GPT-5.5 81.8, ExploitGym 39.5, SEC-bench Pro 69.8) is gated to verified "trusted defenders" under Daybreak β not generally available, not on public API/OpenRouter, no published pricing. Excluded for access, not capability (handled like claude-mythos-5: noted here, NOT given a roster/domain-table row; its evals are cyber-domain only). OpenAI points most orgs to GA "GPT-5.5 + Trusted Access for Cyber + Codex Security". Re-surfaced carried-forward proposals (NOT auto-applied): (1) the now-thrice-recommended opus-4-7βopus-4-8 bump (aloop sleep_build + scripts/cadence/daemon.py:258 + scripts/reports/generate_html_v2.py:54-55) β drop-in, price-neutral, still unapplied; (2) glm-5.1βglm-5.2 on the 2 code-reasoning sites (flows/council-plan/FLOW.yaml:19, api/src/services/improve_cycle.py:68), HOLD the 2 feed-editor sites on glm-5.1; (3) close the 3-review-old cost-tracking gap (add opus-4-8/fable-5/glm-5.2 rows to scripts/inference/costs.py + cost_report.py). Council roster unchanged (opus-4-8/gpt-5.5/gemini-3.1-pro/grok-4.20). Audit: 0 aloop stale (audit roster lacks opus-4-8/glm-5.2 so can't flag those pins β known blind spot); 26 hardcoded "stale" hits all documented residuals or carried-forward bump targets. Est. cost delta β +$1β3/mo. Broad sweep surfaced UNVERIFIED aggregator claims (GPT-5.6/Gemini-3.2/GLM-6; "Fable-5 pulled by export control") β not recorded. Full report: data/reports/2026-06-23-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (gpt-5.5-cyber) |
| 2026-07-05 | Three new models rostered (challengers/pilots) + Fable-5 availability correction. Triggered on kimi-k2.7; broadened after research surfaced two higher-leverage models. (1) claude-sonnet-5 (GA 06-30, $2/$10 intro β $3/$15, perf near Opus 4.8) added as Balanced successor pilot β NOT auto-swapped: a new tokenizer inflates token counts ~1.0-1.35Γ (so "cost-neutral" is intro-only) and voice-critical paths need a drift check; pilot on media_deep_dive first. (2) minimax-m3 (GA 05-31, missed by recent refreshes) added as Budget-MoE successor β same $0.30/$1.20 as m2.7 but 1M ctx + native multimodal; fold into the budget bake-off / pilot on the speculative/* m2.7 sites. (3) kimi-k2.7-code (GA 06-12, the trigger) added as Strong-cheap-code pilot β ~30% fewer thinking tokens but 256K ctx and vendor-only benchmarks; RETAIN k2.6 (its 1M ctx serves the synthesis roles k2.7-code can't) β no current site is a clean swap. New entrants deliberately NOT scored in the 1-10 grid (no verified third-party benchmarks β would be fabrication). Fable-5 correction: the export-control suspension the 06-23 review dismissed as a rumor was REAL (suspended Jun 12-30, restored Jul 1) β slot now carries a geopolitical-availability caveat + needs opus-4-8 fallback for availability, not just refusal. Carried-forward (still unapplied): opus-4-7β4-8 bump now recommended 4Γ (sleep_build aloop + cadence/daemon.py:258 + generate_html_v2.py:54-55) β apply or mark deliberate hold; glm-5.1β5.2 on the 2 code-reasoning sites; the now-4-review-old cost-tracking gap (add opus-4-8/fable-5/glm-5.2/sonnet-5/m3/kimi-k2.7-code rows to costs.py+cost_report.py). Audit: 0 aloop stale by the audit's own roster (known blind spot β it lacks opus-4-8/glm-5.2/sonnet-5/m3/k2.7-code, so it can't flag the sleep_buildβopus-4-7 pin); 27 hardcoded "stale" hits are documented residuals or the carried-forward bump targets. ByteDance Seed 2.1 (Jun 24) + GPT-5.6/Gemini-3.5-Pro (unconfirmed) logged to Open Questions. Full report: data/reports/2026-07-05-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (kimi-k2.7) |
| 2026-07-09 | grok-4.5 lands β new Frontier-agent-cheap pilot tier. SpaceXAI (rebranded from xAI) shipped grok-4.5 (GA 07-08 devs / 07-09 public, $2/$6, cached $0.50, 500K ctx, ~80 TPS, 1.5T "V9" foundation, Cursor-trained). AA-verified: Intelligence Index 54 (#4), Coding Agent Index 76, GDPval-AA v2 1543 Elo. Beats Opus 4.8 on Terminal Bench 2.1 (83.3) + SWE Marathon (29.0); trails on SWE-bench Pro (64.7). Headline = ~2Γ token efficiency (4.2Γ fewer output tokens than Opus 4.8 on SWE-bench Pro) β cost-per-task well under Opus/GPT-5.5. Added to roster + a scored domain mini-table (SCORED unlike the 07-05 entrants β it has verified benchmarks). NOT auto-swapped anywhere. Two disagreements recorded (DeepSWE provider vs neutral harness 62.0/53; Opus-4.8 SWE-bench-Pro 69.2 vs doc's 64.3). Caveats: 500K ctx REGRESSION (grok-4.3 = 1M), no Omniscience β council UNCHANGED (grok-4.20 keeps the honesty seat), modality text-only-confirmed, not-in-EU-til-mid-July, provider rebrand xAIβSpaceXAI (access matrix updated). Routing proposals (NOT auto-applied): (1) pilot grok-4.5 as a challenger on ONE bounded coding path (flows/code-review/synthesize) measuring cost-per-TASK; (2) opus-4-7β4-8 bump now 5Γ-recommended, DECISION FORCED β apply (sleep_build aloop + cadence/daemon.py:258 + generate_html_v2.py:54-55, audit-confirmed still opus-4-7) or mark a deliberate hold; (3) glm-5.1β5.2 on the 2 code-reasoning sites; (4) close the now-5-review-old cost-tracking gap (add opus-4-8/fable-5/glm-5.2/sonnet-5/m3/k2.7-code/grok-4.5 rows to costs.py+cost_report.py); (5) swap the reflect.md:348 placeholder 'kimi-k2.7,qwen-3.5'β'example-model-x' (real-model collision). Audit: aloop_stale empty (known blind spot β roster lacks opus-4-8/glm-5.2/grok-4.5); 210 hardcoded refs, "stale" hits all documented residuals/carried-forward targets; deep-council/FLOW.yaml:28 opus-4-8 flag is a FALSE POSITIVE (correct current ref). Full report: data/reports/2026-07-09-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (grok-4.5) |
| 2026-07-10 | GPT-5.6 family lands (Sol/Terra/Luna) β three SCORED roster rows + council-seat proposal; GPT-Live = watch-only. OpenAI GA'd gpt-5.6 07-09 (preview since 06-26; post-US-Commerce-review β the Fable-5 pattern is now a standing pre-release regime): Sol $5/$30, Terra $2.50/$15, Luna $1/$6; all 1M ctx / 128K out / six effort levels / OpenRouter day-one; new Responses-API programmatic tool calling + subagent spawning + explicit cache breakpoints (writes 1.25Γ input). AA-verified (direct-fetched): Index v4.1 RESCALE β Fable-5 60 #1 / Sol 59 / Terra 55 / Luna 51 (old-index grid numbers NOT comparable; Fable-5 footnote corrected from "~65 single-source" β 60 v4.1 confirmed, partially closing that open question). Sol: #1 Coding Agent Index 80, SOTA Terminal-Bench 2.1 88.8, ALE 53.6 β but SWE-bench Pro 64.6 vs Fable-5's 80 (OpenAI disputes the bench + omits every bench Claude leads) and METR reportedly measured Sol's reward-hacking rate as the highest of any public model it has evaluated (primary unfetched β confirm). Omniscience: hallucination UP vs 5.5 β no council-honesty implications. Luna: Terminal-Bench 84.7 > grok-4.5, but MRCR 41.3 long-context collapse β never >100K ctx. gpt-live-1/-mini (07-08, full-duplex voice, delegates to frontier text model mid-conversation): ChatGPT-only, no API/pricing/OpenRouter β watch item, not rostered (gpt-5.5-cyber handling). No gpt-5.5 deprecation announced, but it's now price-dominated by its successors. Proposals (NOT auto-applied): (1) council OpenAI seat gpt-5.5βgpt-5.6-sol (same $5/$30); (2) opus-4-7β4-8 bump β 6th review, FINAL re-surface: unapplied by next review β auto-marked deliberate-hold; (3) Luna as third arm of the grok-4.5 bounded coding pilot (one pilot, three challengers); (4) fix the audit's frozen snapshot_routing.py:27 CURRENT_ROSTER (root cause of the perpetual "0 aloop stale" lying gauge); (5) cost-tracking gap now 10 models / 6 reviews (adds sol/terra/luna); (6) reflect.md:348 placeholder swap (carry-forward). Audit: aloop_stale [] (blind spot, see #4); 30 hardcoded flags all documented residuals; deep-council:28 opus-4-8 false-positive again. Full report: data/reports/2026-07-10-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (gpt-5.6, gpt-live) |
| 2026-07-18 | Kimi K3 lands β rostered as first open-weight frontier-tier model (WATCH, zero routing changes); opus-4-7β4-8 DELIBERATE HOLD executed. Moonshot K3 (GA 07-16, 2.8T sparse MoE, KDA+AttnRes, $3/$15, cache-hit $0.30, 1M ctx, text+image, single 'max' effort) added to roster + partial domain mini-scores: AA v4.1 = 57 (#3 per coverage / #4 per AA page β discrepancy recorded), $0.94/AA-task, 62 TPS, Arena.ai frontend-code #1 ahead of Fable-5 (independent). Coding suite (Terminal-Bench 2.1 88.3 KimiCode-harness, FrontierSWE 81.2, SWE Marathon 42.0) vendor-only β coding domain deliberately UNSCORED. Pricing regime change: 4Γ k2.6 input β NOT a k2.6 successor β k2.6 retained on every strong-cheap site; K3 competes vs Terra/grok-4.5/Sonnet. Council seat DECLINED (no Omniscience, capacity-limited). Future pilot lane = bounded frontend-generation (generate_html_v2.py / /frontend-slides), gated on weights+license (~07-27), capacity (429s today), and one independent coding number. Executed the 07-10 self-terminating rule: the opus-4-7 pins (aloop sleep_build, cadence/daemon.py:258, generate_html_v2.py:54-55) auto-marked DELIBERATE HOLD β 6 re-surfaces, declined by inaction, reaffirm ~Oct 2026, stops being re-surfaced. Carried forward (4): glm-5.1β5.2 on 2 code-reasoning sites; snapshot_routing.py:27 frozen-roster fix (now missing 11 models incl. K3 β "0 aloop stale" still lies); cost-tracking gap (7 reviews, 11 models, K3 needs a cache-hit column); reflect.md:351 placeholder swap (elevated β this run WAS kimi-triggered). Audit: aloop_stale [] (blind spot unchanged); 30 hardcoded flags all documented residuals/carry-forwards; deep-council:28 false-positive again. Full report: data/reports/2026-07-18-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (kimi-k3) |
| 2026-07-25 | Claude Opus 5 lands β new Frontier-agent LEADER; supersedes opus-4-8 at the same price. Anthropic Opus 5 (GA 07-24, claude-opus-5, $5/$25 β IDENTICAL to opus-4-8) added to roster as the new frontier-lead, with opus-4-8 reframed as superseded-but-retained (the documented cyber-refusal/availability fallback) and Fable-5 narrowed to a niche. AA-verified (direct-fetched): Intelligence Index v4.1 = 61 (max), narrowly #1 of all models (Fable-5 60 / Sol 59 / K3 57); joint-#1 Coding Agent Index (Claude Code @ xhigh); record GDPval-AA v2 1861 + AA-Briefcase 1720 (+146 vs Fable-5); ~26% lower cost-per-task than Fable-5 (at HIGH effort BEATS Fable-5 by 32 Elo at 47% of its cost). Anthropic: OSWorld-2.0 beats Fable-5 at ~1/3 cost, ARC-AGI-3 ~3Γ next-best, +22% hardest agentic coding vs prior Opus. ZDR-clean (no data-retention req, unlike Fable-5); cyber classifiers fire ~85% less than Fable-5 but still refuse offensive-security β opus-4-8 fallback; behind Mythos-5 (Glasswing-only) on cyber. 1M ctx secondary-source (official page stated price only). Added a SCORED domain mini-block (reasoning 10 / coding 10 / long-horizon 10 / cost-eff 7 β AA-verified). Proposals (NOT auto-applied): (1) council Claude deliberation seat + synthesis opus-4-8βopus-5 (setup_models.py + core.py + 5 flows) β SUPERSEDES the 06-17 Fable-5-synthesis pilot (opus-5 = better synthesis at half cost, ZDR-clean, no availability risk); (2) name opus-5 the coding-agent default (doc done; pins follow); (3) add opus-5 $5/$25 row to costs.py+cost_report.py (now 12 models / 8 reviews overdue); (4) fix frozen snapshot_routing.py:27 CURRENT_ROSTER (still opus-4-7, now missing 12 models incl opus-5 β the "0 aloop stale" gauge lies by omission). NEW audit finding: scripts/council/core.py:22 pins opus-4-7, NOT opus-4-8 as this doc claimed β doc/code drift, reconcile to opus-5. DELIBERATE HOLD RESPECTED: the opus-4-7 trio (sleep_build, cadence/daemon.py:258, generate_html_v2.py:54-55) stays held per the 07-18 self-terminating rule β opus-5 is just the new eventual target at the ~Oct-2026 reaffirm; NOT re-surfaced. Est. cost delta β $0 to β$5/mo (all concrete swaps price-neutral; Fable-5 narrowing is a slight savings). No SWE-bench Pro opus-5-vs-Fable-5 head-to-head + no AA Omniscience/TPS pulled β open for next refresh. Full report: data/reports/2026-07-25-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (opus-5) |
| 2026-08-05 | Qwen3.8-Max lands β first Qwen on the roster (WATCH, zero routing changes). Alibaba/Qwen Qwen3.8-Max (GA 08-03, previewed 07-19 at WAIC; 2.4T sparse-MoE, ~95B active, 1M ctx, native text+image+video, $2/$6 + cache-hit $0.25/$0.17, OpenRouter day-one single-provider, open weights ~Aug-10 / license TBD) added to roster as a new Open-frontier-multimodal (watch) tier + a β footnote + provider-matrix row. Independent signal is thin: AA Intelligence Index 53 but from a SEARCH SNIPPET (AA model page 404'd on direct fetch β flagged not-confirmed); on v4.1 that's mid-frontier, BELOW open-weight peer K3 (57). All coding numbers (SWE-bench Pro 67.7, Terminal-Bench 86.6, FrontierSWE 73.5, GPQA-D 92.6) are Alibaba's own table β no independent SWE-bench reproduction. #2 Vision Arena is the one differentiated lane (cheap multimodal + native video @ $2/$6 vs gemini-3.1-pro $2/$12). Deliberately NOT scored in the 1-10 grid (search-snippet AA + vendor-only coding = fabrication risk; same discipline as the 07-05 entrants). China data-jurisdiction/state-law risk β non-sensitive only; self-host once weights land. THREE gates before any pilot: independent coding number + weights/license/capacity + AA vision/OCR number. Competes with grok-4.5 / gpt-5.6-terra / K3 β NOT the cheap band; k2.6/minimax/glm slots untouched. Carried forward (no new info, still bio-zack's call): (1) council opus-4-8βopus-5 Claude seat + synthesis swap + core.py:22 opus-4-7 drift reconcile; (2) council gpt-5.5βgpt-5.6-sol OpenAI seat (gated on METR); (3) cost-tracking gap β now 13 models / 9 reviews overdue, add qwen3.8-max $2/$6 + cache-aware column; (4) frozen snapshot_routing.py:27 CURRENT_ROSTER (still opus-4-7, now missing ~13 models incl opus-5 + qwen3.8-max β "0 aloop stale" lies by omission, 6th review flagged); (5) flows/sleep/prompts/reflect.md:342 placeholder ELEVATED β its literal 'kimi-k2.7,qwen-3.5' example now contains a Qwen name in a Qwen-triggered review = live spurious-detection collision; swap to 'example-model-x'. DELIBERATE HOLD RESPECTED: opus-4-7 trio (sleep_build, cadence/daemon.py:258, generate_html_v2.py:54-55) NOT re-surfaced (07-18 self-terminating rule; reaffirm ~Oct 2026). Audit: aloop_stale [] (known blind spot #4); 218 hardcoded refs, all "stale" hits documented residuals/carry-forwards; deep-council:28 opus-4-8 + model_dialogue.py:30 gpt-5.6-sol are recurring FALSE POSITIVES (correct current refs). Est. cost delta β $0/mo (WATCH model = no routing; council swaps price-neutral; rest is hygiene). Full report: data/reports/2026-08-05-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (qwen3.8-max) |
| 2026-08-13 | Grok 4.6 lands β supersedes grok-4.5 at the same price (SCORED; zero live routing changes). SpaceXAI Grok 4.6 (GA 08-12, x-ai/grok-4.6, $2/$6 base + $4/$12 above 200K, 500K ctx, Feb-1-2026 cutoff, OpenRouter/Cursor/Grok-Build day-one) added to roster as the Frontier-agent-cheap successor; grok-4.5 reframed superseded-but-retained (prefer 4.6 if ever adopted). AA-verified (analysis article direct-fetched; model page NOT fetched β recorded): Intelligence Index 61 (+5 vs 4.5, "matches GPT-5.6 Sol/Sol Max"), GDPval-AA v2 1753 Elo (behind only Opus 5, overlapping CIs w/ Fable-5 + Qwen3.8-Max), AA-Briefcase 1577 Elo (Fable-5-tier), cost-per-task $0.84 (ties K3), ~53-turn/0.5B-input efficiency vs Opus 5's ~103/2.0B. SpaceXAI-harness coding: DeepSWE v1.1 65.9 (+11.9), APEX-Agents 57.5, CursorBench 69.9, FrontierCode 61.3 β no independent SWE-bench Pro/Verified; AA is explicit the lane is agentic + cost-efficiency, NOT raw coding (DeepSWE 65.9 < Sol Max ~73). Added a SCORED domain mini-block (agentic-coding 8 / long-horizon 9 / reasoning 9 / cost-eff 10 β AA-verified, unlike WATCH-only qwen). NEW open flag β AA-index-scale drift: the AA article implies Opus-5 = 63 vs this doc's v4.1 Opus-5 61 / Sol 59 β either AA rescaled or "61=Sol" is a max-effort comparison; unreconciled (needs a direct AA model-page fetch) β do NOT re-pull the reasoning column or trust the parity claim yet. Council UNCHANGED: grok-4.6 evaluated β DECLINED (no AA-Omniscience component; grok-4.20 keeps the honesty seat). Routing proposals (NOT auto-applied): (1) refresh grok-4.5's still-unrun bounded-coding pilot to use grok-4.6 on flows/code-review/synthesize (cost-per-TASK challenger; OFF >200K ctx + OFF council); (2) hygiene bump scripts/experiments/model_dialogue.py:29 x-ai/grok-4.5βgrok-4.6 + scripts/council/core.py:25 registry entry (price-neutral, low-stakes menu refs β grok-4.5 was never a live seat). Carried forward: cost-tracking gap now 14 models / 10 reviews overdue (add grok-4.6 $2/$6 + the 200K-tier band to costs.py+cost_report.py); frozen snapshot_routing.py:27 CURRENT_ROSTER still opus-4-7, now missing ~14 models incl grok-4.6 β the "0 aloop stale" gauge still lies by omission (7th review flagged); flows/sleep/prompts/reflect.md:348 placeholder 'kimi-k2.7,qwen-3.5'β'example-model-x' swap. DELIBERATE HOLD RESPECTED: opus-4-7 trio (sleep_build, cadence/daemon.py:258, generate_html_v2.py:54-55) NOT re-surfaced (07-18 self-terminating rule; reaffirm ~Oct 2026). Audit: aloop_stale [] (known blind spot); 222 hardcoded refs, "stale" hits all documented residuals/carry-forwards; deep-council:28 opus-4-8 + model_dialogue.py:30 gpt-5.6-sol recurring FALSE POSITIVES. Est. cost delta β $0/mo (successor = no live routing; pilot bounded; rest hygiene). Full report: data/reports/2026-08-13-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (grok-4.6) |
| 2026-08-14 | GLM-5.3 + Gemini 3.7 Flash land (same-day double drop) β both rostered WATCH-only; zero routing changes. (1) glm-5.3 (Z.ai, GA 08-14) β a post-training-only refresh of glm-5.2 (SAME 743B base). Vendor-claimed strongest open-weights coder (+50% vs 5.2; DeepSWE 66.9, AutomationBench 48.2, GDPval-AA v2 1769 β all vendor-harness, no independent SWE-bench). The new signal is EMERGENT CYBER: CyberGym 84.5 > Mythos 5 + Sol, 2,436 vulns found across 269 projects β first roster-adjacent option for AUTHORIZED security-testing where every Anthropic frontier model refuses; tracked as a future self-host pilot gated on weights + an independent CyberGym/ExploitBench repro. Per-token price UNPUBLISHED (Coding Plan subs only: Lite $12.6 / Pro $56 / Max $117.6 mo), weights/license ~2wk out (do not assume MIT carries), not on OpenRouter, no AA index β glm-5.2 STAYS the strong-cheap-code pin (no swap); glm-5.2 now superseded (no formal EOL). Added a β roster row + footnote; deliberately UNSCORED in the 1-10 grid (vendor-only = fabrication risk). (2) gemini-3.7-flash (Google, GA 08-13) β cheap agentic/coding/document mid-tier, big gen-over-gen jumps (FrontierCode 34.4β43.6, GDP.pdf 22β34, AutomationBench 17β30.4, all vendor), 1M ctx / 65K out, OpenRouter day-one. vita never rostered 3.5/3.6 Flash, so this is a NEW tier between flash-lite (~$0.10/$0.40, retained) and gemini-3.1-pro ($2/$12) β a candidate cheap-agentic challenger, NOT a flash-lite replacement. PRICING DISAGREEMENT recorded: intro $0.75/$3.75 through 2026-12-31 β $1.50/$7.50 in 2027 (vendor/news) vs OpenRouter $0.375/$1.875 β verify at model card before any volume routing. Added a β roster row + footnote; UNSCORED (no independent benchmark). Council UNCHANGED β neither has an AA-Omniscience component β not a honesty seat (grok-4.20 keeps it). Carried forward (bio-zack's call): frozen snapshot_routing.py:27 CURRENT_ROSTER still opus-4-7, now missing ~15 models incl opus-5/grok-4.6/glm-5.3/gemini-3.7-flash β the "0 aloop stale" gauge still lies by omission (8th review flagged; single highest-leverage hygiene fix); cost-tracking gap now ~15 models / 11 reviews overdue (add opus-5/grok-4.6/gemini-3.7-flash rows, glm-5.3 when priced, to costs.py+cost_report.py); flows/sleep/prompts/reflect.md:348 placeholder 'kimi-k2.7,qwen-3.5'β'example-model-x'; council opus-4-8βopus-5 + core.py:22 opus-4-7 drift; gpt-5.5βgpt-5.6-sol gated on METR. DELIBERATE HOLD RESPECTED: opus-4-7 trio NOT re-surfaced (07-18 rule; reaffirm ~Oct 2026). Audit: aloop_stale [] (known blind spot β frozen roster); 222 hardcoded refs, "stale" hits all documented residuals/carry-forwards; deep-council:28 opus-4-8 + model_dialogue.py:30 gpt-5.6-sol recurring FALSE POSITIVES. Est. cost delta β $0/mo (both WATCH = no routing; rest hygiene). Full report: data/reports/2026-08-14-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (glm-5.3, gemini-3.7-flash) |
| 2026-09-03 | Claude Fable 5.1 + Mythos 5.1 land β fable-5-1 supersedes fable-5 as Frontier-max leader; zero live routing changes. Anthropic (GA 09-01) claude-fable-5-1: SAME per-token $10/$50 as fable-5 but ~25% cheaper EFFECTIVE (up to ~45% agentic) via token-efficiency + a 75% cache-read cut ($1.00β$0.25), batch $5/$25, 1M ctx/128K out. AA-independent: Intelligence Index (v4.1.1/"2026" scale) 65.7 β new #1 (Opus 5 63.0 / Fable 5 62.1). Anthropic table: Terminal-Bench-Science 52.6 (Opus 5 29.0), GDPval-AA v2 1853, OSWorld 41.7, AutomationBench 31.4, CursorBench 73.4. Updated roster (fable-5-1 leader row + superseded fable-5 row), β‘β‘ footnote, provider matrix, "hardest builds" guidance, and added a SCORED domain mini-block (it has an AA-independent index, unlike WATCH entries β reasoning/coding/long-horizon 10, cost-eff 3β4). Three genuinely-new items: (1) CYBER β fable-5-1 can DISCOVER vulns (not exploits), ~60% fewer Claude Code cyber-safeguard interventions β first GA/reachable Anthropic option for authorized/defensive security-testing (whole roster refused; glm-5.3 vendor-only+no-weights, Mythos 5.1 trusted-access); (2) invisible EU-AI-Act watermark on fable-5.1/mythos-5.1 text/file output (Anthropic-detection-API-only) β new consideration for anything published as bio-zack; (3) AA re-scaled to v4.1.1 β RESOLVES the 08-13 Opus-5=63 open flag (AA did rescale) and drops Sol (58.9) below Fable 5 + Opus 5. Mythos 5.1 = same model, cyber/bio safeguards, trusted-access US-orgs-only β NOT vita-adoptable. Council UNCHANGED (fable-5-1 no Omniscience component; opus-5 still cheaper+ZDR-clean+un-watermarked β stays default + synthesis pick; the opus-4-8βopus-5 seat-swap + gpt-5.5βsol proposals still stand). TOP recommendation β the audit known_roster proves the frozen snapshot_routing.py:27 CURRENT_ROSTER STILL never updated (lists opus-4-7/glm-5.1/kimi-k2.6, missing ~16 live models incl opus-5/fable-5-1); aloop_stale = [] because the gauge is BLIND, not clean β 9th consecutive flag, escalate + add a deliberate-hold allowlist for the opus-4-7 pins so the fix doesn't re-nag the 6Γ-declined bump. One genuine live stale ref: scripts/seo/geo.py:41 claude-sonnet-4-5βclaude-sonnet-4-6 (+ confirm the geo.py:40 gemini-3.5-flash GEO-panel ref). New leaderboard-only WATCH sightings: Gemini 3.8 Flash (58.7), GLM-5.3-Flash (57.5), Meta Muse Spark 1.2 (56.8 β first Meta), Qwen3.8-Flash-Next (55.8); glm-5.3 now AA-scored (59.5). Carried forward (bio-zack's call): cost-tracking gap now ~16 models / 12 reviews (add fable-5-1/opus-5/grok-4.6 rows to costs.py+cost_report.py); reflect.md:348 placeholder swap; council opus-4-8βopus-5 + core.py:22 opus-4-7 drift; gpt-5.5βsol gated on METR. DELIBERATE HOLD RESPECTED: opus-4-7 trio NOT re-surfaced (07-18 rule; reaffirm ~Oct 2026). Audit: aloop_stale [] (blind spot β see TOP rec); 225 hardcoded refs, "stale" hits all documented residuals/carry-forwards; deep-council:28 opus-4-8 + model_dialogue.py:30/core.py:23 gpt-5.6-sol recurring FALSE POSITIVES (correct current refs; matcher mis-flags 5.6 as "gpt-5 base"). Est. cost delta β $0/mo. Full report: data/reports/2026-09-03-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (claude-fable-5-1, claude-mythos-5-1) |
| 2026-09-04 | GPT-6 Astra lands β rostered WATCH (SCORED, has AA-independent numbers); zero live routing changes. OpenAI (GA 09-03, gpt-6-astra) β first 6th-gen model, takes GPT-5.6 Sol's top slot. $10/$50, cached $1, cache-write $12.50, >272K prompt rebills whole request to ~$20/$75, 1.05M ctx / 128K out, cutoff 2026-04-30, effort lowβmax, NOT on OpenRouter (OpenAI-API-only; Daybreak-cohort-first). AA-independent (article direct-fetched): Intelligence Index (v4.1.1) 61 (max) β TIES Sol + Grok-4.6, 5 below Fable 5.1 (66); AA Coding Agent Index 67 β Opus 5; GDPval-AA v2 β~80 Elo vs Sol (a DROP); 51% hallucination at max effort. Genuine leaps are all VENDOR-harness: OSWorld 2.0 72.6 (~2Γ Sol's speed β computer-use SOTA), Terminal-Bench 4.0 57.7 (> Fable 5.1 55.8), and CYBER (first OpenAI model rated CRITICAL under the Preparedness Framework; ExploitBench 100% w/o safeguards). Added roster row (Frontier-6th-gen watch) + β
footnote + provider-matrix note + a SCORED domain mini-block (reasoning 9 / coding 9 / computer-use 10 / cost-eff 3). Two flags: (1) the "61 = Sol" tie is AA's MAX-effort read β do NOT overwrite the doc's v4.1.1 baseline (Sol 58.9 / Fable 5.1 65.7 default-blend); (2) the cyber headline is a TRAP β exploit-CREATION is GATED behind Daybreak + BLOCKED in the production API, so a normal API caller still gets a refuse (like the rest of the roster minus Fable 5.1's discover-not-exploit GA path); Astra is NOT vita's authorized-security option. Routing: WATCH, zero pin moves β $10/$50 = 2Γ Opus 5 / 2Γ Sol for equal-or-lower measured intelligence, no adopt rationale. Council: evaluated β DECLINED (no Omniscience, $10/$50, 51% hallucination β grok-4.20 keeps the honesty seat). TOP recommendation (10th consecutive flag): aloop_stale=[] is the frozen snapshot_routing.py:27 CURRENT_ROSTER BLIND SPOT β the audit known_roster still lists opus-4-7/glm-5.1/kimi-k2.6/gpt-5.5, missing ~17 live models incl opus-5/fable-5-1/grok-4.6/astra, so the gauge cannot see a stale pin; refresh it + add a deliberate-hold allowlist for the opus-4-7 pins so the fix doesn't re-nag the 6Γ-declined bump. AUDIT POSITIVE: scripts/council/core.py:22/33 now correctly pin claude-opus-5 + :23 gpt-5.6-sol β the 07-25 opus-4-7 drift is RESOLVED in core.py (open: confirm flows/council/setup_models.py matches; if it still shows opus-4-8/gpt-5.5 the standing seat swaps remain pending on that path). One genuine live stale ref: scripts/seo/geo.py:41 claude-sonnet-4-5βclaude-sonnet-4-6 (carried from 09-03, still unfixed). Carried forward (bio-zack's call): cost-tracking gap now ~17 models / 13 reviews (add gpt-6-astra $10/$50 + the >272K rebill tier + the still-missing opus-5/fable-5-1/grok-4.6 rows to costs.py+cost_report.py); flows/sleep/prompts/reflect.md:348 placeholder 'kimi-k2.7,qwen-3.5'β'example-model-x'. DELIBERATE HOLD RESPECTED: opus-4-7 trio (sleep_build, cadence/daemon.py:258, generate_html_v2.py:54-55) NOT re-surfaced (07-18 rule) β note the ~Oct-2026 quarterly reaffirm window is now imminent. Astra ZDR/watermark/vision/independent-SWE-bench NOT researched this run β pull before any routing. Audit: aloop_stale [] (blind spot β see TOP rec); 225 hardcoded refs, "stale" hits all documented residuals/carry-forwards except the one geo.py live stale; deep-council:28 opus-4-8 + model_dialogue.py:30/core.py:23 gpt-5.6-sol recurring FALSE POSITIVES (correct current refs; matcher mis-flags 5.6 as "gpt-5 base"). Est. cost delta β $0/mo. Full report: data/reports/2026-09-04-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (gpt-6-astra) |
| 2026-09-21 | Trigger was a PHANTOM β claude-opus-5-5 does not exist; NOT rostered. Real changes found while verifying: Grok 4.7 rostered (supersedes grok-4.6), Sonnet 5 price corrected to permanent $2/$10, Muse Spark 1.3 rostered WATCH, AA index re-scaled (v4.3.2), council setup_models.py drift confirmed, first near-term retirement floors. The opus-5-5 detection came from a single unverified X post (codename "claude-wafer-eap", claimed 9-22 release) plus a 20%-off "price sheet" and a demo video retracted as recycled footage β verified ABSENT at platform.claude.com models/deprecations tables, anthropic.com/news, openrouter.ai/anthropic, AA. Re-check 9-23; if real, re-run from the announcement. (1) grok-4.7 (SpaceXAI, GA 09-21, same $2/$6 + 500K as 4.6; AA v4.3.2 = 46 xhigh, $3.74/task, 38.8 TPS, "slow + very verbose"; vendor: CursorBench 4.0 46.3 > Sol 41.7, DeepSWE 71.0, Terminal-Bench 4.0 38.0) β new Frontier-agent-cheap row, grok-4.6 superseded, β footnote, partial mini-scores (cost-eff held at 8, NOT inheriting 4.6's 10 β effective-cost headline undemonstrated). (2) sonnet-5 $2/$10 is PERMANENT (pricing page: the 9-01 rise to $3/$15 "will not occur"); official tokenizer inflation ~30% β cost-neutral vs 4.6. Roster row + ΒΆΒΆ corrected. (3) muse-spark-1.3 (Meta, GA 09-02, AA v4.3.2 = 48 > Sol 47, $1.60/task cheapest top-12; $1.25/$4.25 snippet-only; contributor tier trains on data) β WATCH row, zero routing. (4) AA Index v4.3 (09-07) / v4.3.2: Terminal-Bench 4.0 + AutomationBench-AA replace 2.1 + ΟΒ³-Banking; new top: Fable 5.1 53 = Astra 53 > Opus 5 51 > Muse 48 > Sol 47 > Grok 4.7 46 β all v4.1.1 numbers in the doc flagged retired; Astra now a TIE with Fable 5.1 at ~43% of its cost-per-task. (5) Council drift CONFIRMED: flows/council/setup_models.py:5-8 = opus-4-7 / gpt-5.5 / gemini / grok-4.20 (this doc claimed opus-4-8 since 04-23) vs core.py = opus-5 / sol / gemini / grok-4.5 / k3 β bump :5 to opus-5 (price-neutral). (6) Retirement floors: sonnet-4-5 9-29, haiku-4-5 10-15 (no notices yet; 60-day rule) β scripts/seo/geo.py:41 sonnet-4-5 pin is now time-sensitive (β sonnet-5); haiku pins at feed/constitutions/extractor.py:34, media/bookmarks.py:148, flows/research-proposal/FLOW.yaml:41 need a named successor (no Haiku 5). 7 proposals (NOT auto-applied): hold opus-5-5 as rumor; geo.py:41 β sonnet-5 (+ :40 gemini-3.5-flash check); setup_models.py:5 β opus-5; sonnet-5 pilot on media_deep_dive + epd47_generate (now cost-flat, was cost-worse); grok-4.5β4.7 menu refs (core.py:25, model_dialogue.py:29) + pilot target; haiku successor decision; frozen snapshot_routing.py:27 roster (11th flag). Council: keep core.py; grok-4.7 + muse DECLINED (no Omniscience). DELIBERATE HOLD RESPECTED: opus-4-7 trio NOT re-surfaced (~Oct reaffirm imminent). Audit: aloop_stale [] (blind spot); 227 refs / 35 flagged, 2 genuine live stale (geo.py:41, setup_models.py:5 β the latter invisible to the audit). Est. cost delta β $0 to β$2/mo. Full report: data/reports/2026-09-21-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (claude-opus-5-5 β rumor, verified non-existent) |
| 2026-09-21 | bio-zack approved the 09-21 review wholesale ("please make all these recommended changes", by email) β all 7 recommendations APPLIED the same evening; first review in this series to be executed rather than filed. (1) opus-5-5 NOT rostered; a primary-source re-check is queued for 9-23 9am (proactive.queue id 12657, trigger model_guidance_recheck) instead of relying on the next run to remember. (2) scripts/seo/geo.py:41 sonnet-4-5 β sonnet-5, ahead of the 9-29 retirement floor; DataForSEO's /ai_optimization/claude/llm_responses/models endpoint was queried first to confirm it serves claude-sonnet-5 (it does β that panel takes DataForSEO's model names, not Anthropic's). (3) flows/council/setup_models.py:5 opus-4-7 β opus-5, price-neutral, ends five months of doc drift. (4) Sonnet 5 pilot STARTED on media_deep_dive + epd47_generate; voice paths untouched. (5) Haiku successor decided per-site β feed/constitutions/extractor.py β sonnet-5 (identity-adjacent, same vendor, same prompt shape) with thinking:{type:disabled} added, because sonnet-5 thinks by default and adaptive thinking would have eaten its max_tokens=400 and truncated the JSON; media/bookmarks.py β gemini-3.1-flash-lite (already on OpenRouter, 10Γ cheaper, matches the media triage tier); flows/research-proposal β sonnet-5 (a 4-vendor panel β flash-lite would have duplicated the Google arm). (6) grok-4.5 β grok-4.7 at core.py:25 + model_dialogue.py:29, cost comment resynced, and the still-unrun bounded-coding pilot retargeted to 4.7. (7) snapshot_routing.py rebuilt roster-driven β see the Audit snapshot note; it exposed and fixed a truncating GREP_PATTERN as a side effect. |
bio-zack's approval of the 2026-09-21 review |
| 2026-09-23 | The 09-21 "PHANTOM" call was WRONG β claude-opus-5-5 shipped 09-22 and is the new Frontier-lead, superseding opus-5 at a 20% LOWER price. Plus a same-day OpenAI price war (GPT-6 Sol + Luna) and the first genuinely new audit finding since the roster fix. The queued 9-23 re-check (proactive.queue id 12657) fired and found the model live at every primary source. Every number in the "leaked price sheet" the 09-21 review refused to trust was exact ($4/$20, cache-read $0.20, cache-write $5). Withholding the roster row on 09-21 was correct; asserting the model did not exist was the overclaim β second instance after the Jun-2026 Fable-5 export-control suspension (dismissed 06-23, real 07-05). Standing lesson, now with two data points: quarantine rumors, never assert absence from them. (1) claude-opus-5-5 (GA 09-22) β $4/$20, cache read $0.20 (0.05Γ β a third distinct cache shape), cache write $5/$8, batch $2/$10, fast $8/$40, 1M ctx / 128K out, cutoff Jun 2026, effort lowβmax default medium, floor 2027-09-22, OpenRouter day-one. AA v4.3.2 = 58 (max), #1 of 212, +5 over Fable 5.1 and Astra (both 53); leads 6 of 10 components β HLE 61.4, SciCode 66.9, GDPval-AA v2.1 1846 (+111/+138), AA-Briefcase 1822 (+143), AA-Omniscience, AutomationBench-AA; behind on CritPt/AA-LCR/GDP.pdf. Vendor: Terminal-Bench 4.0 66.4, CursorBench 4.0 57.8, OSWorld 2.0 81.8, Chartography 89.0. THE COST TRAP recorded prominently: ~1.6Γ verbosity (260M output tokens in AA's eval vs 88M median) means AA's cost-per-task is FLAT β $5.98 vs opus-5's $5.86. Adopt for quality, not savings; the real win is on cache-heavy/short-output paths. Disagreement flagged: Terminal-Bench 4.0 vendor 66.4 (Astra 57.9) vs AA 59.6, a TIE with Astra β different harnesses, both recorded. ZDR available; now watermarked (opus-5 was the last un-watermarked frontier Claude); thinking cannot be disabled; cyber redirects to opus-4-8; new anti-distillation safeguard for API accounts created after 08-31. Added roster row + β footnote + a SCORED domain mini-block (honesty 8β9, the first movement on that row in the doc's history; cost-eff capped at 8 for the verbosity). (2) GPT-6 Sol + Luna shipped the SAME DAY β Sol $2/$10 (AA v4.3.2 48, $0.13β$1.06/task, 872K ctx), Luna $0.10/$0.50 (~10Γ under gpt-5.6-luna). Vendor-vs-independent gap flagged: OpenAI's "beats Opus 5 at 9% of cost" is one benchmark; AA's composite has Sol 48 < Opus 5 51 < Opus 5.5 58. This makes the standing 07-10 gpt-5.5βgpt-5.6-sol council proposal obsolete β retarget to gpt-6-sol. Both rostered + β¦β¦ footnote; Luna WATCH pending an MRCR re-test (its predecessor collapsed at 41.3). (3) deepseek-v4.1-flash (09-10) rostered WATCH + ββ footnote β time-of-day pricing, a roster first: off-peak $0.15/$0.60, peak DOUBLE, window MonβFri 01:00-04:00 + 06:00-10:00 UTC, which is exactly when vita's nightly rollup (07:05 UTC) and /sleep (07:30 UTC) run. 2.1β4.3Γ the incumbent v4-flash's output β not a cheaper successor. (4) No new deprecations, but opus-5 moved to "Legacy (still available)" on models-overview while the deprecation table still says Active/2027-07-24 β both recorded. sonnet-4-5 (floor 09-29, six days out) and haiku-4-5 (10-15) still have no notice β 60-day rule pushes failure to ~late Nov. Sonnet 5.5 + Haiku 5.5 signalled "within weeks" β let the running sonnet-5 pilot ride rather than widening it. (5) AUDIT β first genuinely new finding since the 09-21 roster fix: six downstream council flows are TWO GENERATIONS stale on opus-4-7/opus-4-8, and this doc has promised that sync since 07-25 without ever applying it (simple-council:40, council-review:38, vision-strategy:169, codebase-research:172 = opus-4-7; deep-council:28, fusion-council/setup_models.py:16 = opus-4.8). The 09-21 pass fixed only flows/council/setup_models.py:5. None is under the opus-4-7 deliberate hold. aloop_stale is legitimately [] this run (all 25 pins check out; sleep_build correctly bucketed held). 6 proposals (NOT auto-applied): (a) council Claude seat + synthesis opus-5 β opus-5-5 (core.py:22+:33, setup_models.py:5); (b) sweep the six downstream council flows β opus-5-5; (c) OpenAI seat gpt-5.5 β gpt-6-sol (retargets the obsolete 5.6-sol proposal; METR flag un-rechecked for the new model); (d) re-examine the $20/$60 grok-4.20 honesty seat β first time any model has cleared the AA-Omniscience gate, but AA's framing is that Anthropic leads on calibration and xAI on accuracy, so this is NOT an automatic retirement; pull same-scale numbers first; (e) cost-tracking rows for opus-5-5/gpt-6-sol/gpt-6-luna with a per-model cache-multiplier column (0.1Γ standard / 0.05Γ opus-5-5 / 0.025Γ Fable β a flat column mis-costs all three); (f) doc/roster update (applied here). DELIBERATE HOLD RESPECTED: opus-4-7 trio NOT re-surfaced β the ~Oct-2026 quarterly reaffirm opens next month and opus-5-5 is the new eventual target there. Est. cost delta β β$4 to β$10/mo (per-token cuts on low-volume council paths, partly offset by opus-5-5's verbosity). Full report: data/reports/2026-09-23-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (claude-opus-5-5 β REAL, confirmed at primary sources) |
| 2026-09-29 | Claude Sonnet 5.5 lands β new Balanced LEADER at sonnet-5's exact price; Haiku 5.5 announced but NOT shipped; the audit's clean bill of health is wrong again, for a new reason. (1) claude-sonnet-5-5 (GA 09-28) β $2/$10 IDENTICAL to sonnet-5, cache read $0.20 / write $2.50, 1M ctx / 128K out, cutoff Jun 2026, adaptive thinking, effort lowβmax default high (higher than opus-5-5's medium), floor 2027-09-28, ZDR-clean, OpenRouter day-one (5 providers). AA v4.3.2 = 56 (max), #3 of 216 β sonnet-5 is 38 (#58): an 18-point jump for zero per-token cost, the largest same-price generational move this doc has recorded; xhigh 52, high 47, 137.6 TPS (fastest Sonnet). Vendor table beats opus-5-5 on Terminal-Bench 4.0 (70.6 vs 66.4) and near-ties on GDPval 1844/1846, AA-Briefcase 1811/1822, OSWorld 2.1 80.1/81.8, CursorBench 55.5/57.8. Added roster row + β footnote + a SCORED domain mini-block; sonnet-5 reframed superseded-but-retained (it is now sonnet-5-5's cyber fallback target, and the two pins that can't take 5.5 stay on it). THE COST TRAP, in OPPOSITE directions this time β both recorded: Anthropic claims "up to 30% LESS per task" (Balyasny 121K tokens/answer vs 497K; Base44 3.6 iterations vs 7.7); AA measures $7.60/task vs sonnet-5's $5.09 β ~49% MORE, and above opus-5-5's $5.98 β "notably fast, however very verbose", 410M output tokens in eval. Not reconcilable from outside (AA = fixed hard-reasoning at max effort; Anthropic = iteration counts on real work); neither stands alone. And AA's effort frontier has opus-5-5 at xhigh tying it at 56 for $3.46, so sonnet-5-5-at-max is dominated by the flagship at a lower tier β cost-efficiency scored 7, BELOW opus-5-5's 8, despite the cheaper sticker. Adopt for capability + speed, not savings. (2) claude-haiku-5-5 DID NOT SHIP β Anthropic's own page commits it to "the coming weeks" with no ID/date/price/benchmark; the platform table still lists Haiku 4.5 as the only Haiku. Rostered with null specs + ββ footnote, deliberately UNSCORED. The 09-21 phantom lesson in its other direction: a primary source exists, so existence is FACT while specs are UNKNOWN. Do not extrapolate haiku-4-5's $1/$5 β sonnet-5-5 held price while opus-5-5 cut 20%; there is no generation rule. (3) ROUTING β 3 clean swaps, 2 deliberate HOLDS, and the holds are the real finding. sonnet-5 is live at five sites; only three are one-line edits. scripts/feed/constitutions/extractor.py:34+72 HELD: it passes thinking={"type":"disabled"}, which adaptive-thinking Claudes 400 on (fable-5/5.1 both; opus-5-5 can't disable thinking) β and deleting the param re-opens the exact bug the 09-21 pass fixed (adaptive thinking eats max_tokens=400, truncating the JSON). Needs a max_tokens raise and param removal and a parse check β a code change, not a swap. scripts/seo/geo.py:44 HELD: those are DataForSEO model_name values, not vendor API ids (file comment gates changes on their models endpoint), and the file deliberately freezes versions for run-over-run comparability; the 09-21 bump was forced by sonnet-4-5's floor, and sonnet-5's floor is 2027-06-30 β no forcing function, so bumping would trade comparability for nothing. (4) AUDIT β aloop_stale=[] is WRONG, for a NEW reason. The 09-21 structural fix holds, but the roster has fallen behind by maintenance lag: known_roster is missing opus-5-5/sonnet-5-5/gpt-6-sol/gpt-6-luna/deepseek-v4.1-flash, and superseded_map is a generation stale (sonnet-4.6βsonnet-5 should be βsonnet-5-5; opus-4.8/4.7βopus-5 should be βopus-5-5; gpt-5.5βgpt-5.6-sol should be βgpt-6-sol). Because claude-sonnet-5 is in the roster and absent from the supersession map, both live pilot pins read CLEAN while being stale β the audit could not surface this run's headline rec. A zero is not evidence until the filter is proven able to return something. Distinct from the 09-21 bug (that was structural; this is a process with no trigger linking roster additions to roster updates). (5) TWO LONG-CARRIED ITEMS CLOSED, both verified in code rather than assumed: all six downstream council flows are now on opus-5-5 + gpt-6-sol β a sync this doc promised on 07-25 and carried unapplied for nine weeks β so all eight council entry points agree for the first time in the doc's history; and the cost-tracking gap, carried 13 reviews, is CLOSED (costs.py:27-38 holds opus-5-5/opus-5/fable-5-1/sonnet-5 with per-model cache multipliers and a "never derive cache_read" comment). Only claude-sonnet-5-5 needs adding (rec 4, identical rates to sonnet-5). (6) No new deprecations. sonnet-4-5's floor passed today (09-29) and haiku-4-5's is 10-15, both notice-free β 60-day rule pushes failure to ~late Nov. haiku-4-5 is a NON-ISSUE for vita: zero live code pins remain (routed away 09-21; only data/tokens/*.json history, catalog by design). Quieter change: Anthropic moved sonnet-5, sonnet-4-6 AND sonnet-4-5 into the overview page's Legacy list while the deprecation table calls them Active β so vita's live balanced default (sonnet-4-6, 26 refs) and its running pilot are both off the recommended list. Council UNCHANGED β KEEP: sonnet-5-5 evaluated β DECLINED (56 < opus-5-5's 58, no Omniscience lead, council volume too low for the per-token delta to matter); grok-4.20 keeps the honesty seat. 6 proposals (NOT auto-applied): (a) .aloop/config.json:39+:90 sonnet-5 β sonnet-5-5 (re-point the pilot, don't widen it) + watch tokens; (b) flows/research-proposal/FLOW.yaml:41 β sonnet-5-5; (c) HOLD extractor.py + geo.py with the rationale above; (d) add claude-sonnet-5-5 to costs.py; (e) refresh snapshot_routing.py roster + supersession map; (f) keep the 26 sonnet-4-6 voice refs held one more cycle pending a drift check. DELIBERATE HOLD RESPECTED: the opus-4-7 trio (sleep_build, cadence/daemon.py:258, generate_html_v2.py:54-55) NOT re-surfaced β the ~Oct-2026 quarterly reaffirm window opens in 2 days; opus-5-5 is the eventual target there. Est. cost delta β $0 to +$3/mo (all swaps per-token identical; direction depends on which verbosity read holds). Full report: data/reports/2026-09-29-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (claude-sonnet-5-5 β shipped; claude-haiku-5-5 β announced, not shipped) |
| 2026-10-02 | Gemini 4 Argon lands and is UNROUTABLE; the urgent finding is a 21-day shutdown date; and the trigger model was the third-most-important thing in the run. (1) gemini-4-argon (Google, 09-30) rostered WATCH with zero routing and no API id β AA v4.3.2 53 (high), #8/224, +23 over gemini-3.1-pro-preview, back in the top three labs, #1 AutomationBench-AA (78%), 1M OUTPUT tokens (from 64K), 15% Omniscience hallucination (lowest above index 45), shipped to Fairwind cyber defenders without cyber guardrails. But: no published API model id, absent from OpenRouter/Vertex/CLI/Cursor/Copilot, Fairwind-only with no GA date; $2/$10 is an undated promotion reverting to $4/$20 ($1.99 β $3.98/task, ~22% above Astra); 62k output tokens/task vs Astra's 27k; and the hallucination headline sits on 50% accuracy vs Astra's 63% β it abstains more, it does not know more. Google's "significantly higher than Astra/Fable/Opus" (Vals) does not survive AA, which has it tied with Astra and five behind opus-5-5. PARTIALLY scored (agentic 10, reasoning 9, coding 7, cost-eff 5, honesty 8). (2) MISS CAUGHT: gpt-6.1-sol shipped at DevDay 09-29, one day before the last refresh, and was not caught. Rostered as the new Frontier-value leader, superseding gpt-6-sol at the identical $2/$10: AA 48 β 52, cost/task $1.05 β $0.72, and its high tier (50 @ $0.32) is cheaper per task than any tier of any model in AA's top 25 β ~90% of opus-5-5's score at ~12% of its per-task cost. Scored cost-efficiency 10, the first non-budget model to earn it; every capability number is OpenAI's own harness and openai.com 403'd, so coding/computer-use left unscored and the gpt-5.6-luna MRCR 41.3 collapse carried forward as a hard long-context caveat. (3) HARD DEADLINE, 21 DAYS: gpt-4.1-nano shuts down 2026-10-23 and is live as BUDGET_MODEL at scripts/photos/extract.py:47 β proposed google/gemini-3.1-flash-lite (cost-neutral, already the budget-fast pin, no new vendor surface), plus the live MODEL_PRICING dict at :51-53. Nothing else on the board breaks if ignored. Also newly dated: gpt-4o-mini-transcribe removed 2027-02-27 (api/src/services/agenda_realtime.py:30). (4) 12 FALSE POSITIVES CLEARED and a two-review caveat CLOSED: gpt-4o-mini and gpt-4.1-mini are NOT deprecated at OpenAI β verified at the deprecations page today, where the 09-23 and 09-29 reviews had both deferred with "verify before sweeping". Do not sweep the 11 gumball/photos refs. (5) FIRST ACTIONABLE DEPRECATION IN THE DOC'S HISTORY: claude-sonnet-4-5-20250929 deprecated 09-30, retires 2026-11-30, replacement sonnet-5-5 β landing the day after the last refresh and vindicating four reviews' "nothing fails before late Nov" call. Blast radius NIL (zero live routing pins; geo.py already reads sonnet-5; the only hits are historical price rows that must stay). haiku-4-5's floor is 13 days out, still no notice. haiku-5-5 re-verified absent from the platform table. (6) ZERO of the five 09-29 recommendations were applied in three days β re-asserted as one batch with no new argument (config.json:39/:90 + research-proposal:41 sonnet-5 β sonnet-5-5, costs.py sonnet-5-5 row). (7) Audit: aloop_stale=[] wrong a SECOND consecutive run for the same unfixed reason (snapshot_routing.py:29-73 + :77-87 roster/map maintenance lag β still a data edit, still not made), plus a new double-count (.aloop/config.json:71 is both held and stale because DELIBERATE_HOLDS matches mode-name-or-path and the config path is neither). (8) The hold trio is now a PAIR β scripts/reports/generate_html_v2.py no longer exists; the opus-4-7 reaffirm window opens this month, eventual target now opus-5-5. (9) The research-proposal 4-vendor panel is stale in THREE of four arms (gpt-4.1-mini / gemini-2.5-flash-preview-05-20 / deepseek-r1), which the 09-29 review missed by reading only the Anthropic arm. Council: ONE seat changes (OpenAI β gpt-6.1-sol at both entry points); argon declined (unroutable, and 53 < 58); grok-4.20 keeps the honesty seat. Est. cost delta β $0/mo (β$1 to +$3) β honestly not a savings run; the one real lever sits on paths vita does not route to yet. Full report: data/reports/2026-10-02-model-guidance-review.md. |
flows/model-guidance-review automated run, trigger: new-model-detected (gemini-4-argon) |
Methodology β how to refresh
When to refresh:
- A new frontier model drops (GPT, Claude, Gemini majors)
- A new cheap-strong challenger lands (DeepSeek, Kimi, GLM, MiniMax, Qwen majors)
- Existing model prices change meaningfully (>20%)
- A vita workload shows quality issues traceable to model choice
- Quarterly, even with no new models, to catch drift
How the automated flow works (flows/model-guidance-review):
- Snapshot current routing. Reads
.aloop/config.json,flows/council/setup_models.py, and greps vita/gumball/cadence for hardcoded model IDs. Writesdata/model-guidance/routing-audit-<date>.json. - Research new models. Agent with Camofox + WebSearch pulls provider announcements, pricing pages (
artificialanalysis.ai,pricepertoken.com,llm-stats.com, vendor docs), and top independent benchmarks (SWE-bench Pro, AA Intelligence Index, Terminal-Bench 2, Omniscience). - Score by domain. Agent updates the domain table above using the published numbers + vita-calibrated weights (tool use stability and voice fidelity matter more than raw reasoning for our workloads).
- Generate diff. Agent proposes concrete changes to
.aloop/config.jsonand councilsetup_models.py. Proposals are diffs, not auto-applied. - Email bio-zack. HTML summary with: new models, recommendation summary, diff, and a link to this doc for deep-dive.
- Persist. Updates this doc (
docs/pages/model-guidance.md), appends to change log, and writes a signed report todata/reports/YYYY-MM-DD-model-guidance-review.md.
Manual refresh:
# from vita root
stepwise run flows/model-guidance-review --name model-review --wait
Sleep-triggered refresh:
The /sleep reflect step watches for new-model mentions in data/discoveries/*.json (from web_discover), in recent bio-zack telegram/voice mentions, and in data/reports/*. If any named model is not covered in this doc's roster, the refresh job is enqueued.
Calibration notes:
- vita-specific weights diverge from public benchmarks in three places: tool use stability (we run long agent sessions, partial-failure shape matters), voice fidelity (identity coherence is a first-class requirement), and streaming speed (telegram sessions are latency-sensitive).
- Non-hallucination matters disproportionately for the coaching + doctor agents β grok-4.20's Omniscience lead is why it gets a council seat.
- OpenRouter adds ~5% markup + latency; direct SDK preferred for high-volume modes.
- When in doubt: cheaper model first, escalate only on measured quality regression. VITA #1 issue means bad outputs persist β but so does over-spending.
Open questions
- gemini-4-argon access watch (new 2026-10-02): the model is rostered on evidence but cannot be pinned β no API model id exists at any primary source. Three things must land before any routing conversation: (a) a published API model id (and until one exists, treat any id seen in the wild as a rumor β the 09-21 phantom lesson cuts both ways: do not assert absence from a rumor, and do not pin from one either); (b) confirmation of the steady-state price β the $2/$10 is explicitly introductory with no stated end date, and the whole cost case inverts at $4/$20; (c) GA beyond Fairwind. Also unstated and worth pulling when it ships: knowledge cutoff, the effort-tier set (AA benchmarks a "high" tier so tiers exist), SynthID watermark status (relevant to anything published as bio-zack, same consideration logged for fable-5-1 and opus-5-5), and ZDR posture. If it reaches GA with the cyber posture intact it becomes the strongest authorized-security-testing candidate on the roster β ahead of Fable 5.1's discover-not-exploit path and glm-5.3's unreproduced CyberGym β which is the one lane where it would be worth real integration work rather than a price comparison.
- gpt-6.1-sol β the cost lever vita cannot currently pull (new 2026-10-02): on the one independently-measured axis it is the best value on the board (50 @ $0.32/task), and vita has no path that routes there. The council seat is cents. The 9 research/synthesis modes are on deepseek-v4-flash at $0.14/$0.28, which is far cheaper per token β so this is not a swap, and pretending otherwise would be the "cheaper per token β cheaper per task" trap running in reverse. The honest question for bio-zack: is there a workload that currently escalates to opus-5-5 or sonnet-5-5 where 52-at-$0.72 would do? If yes, that is the single largest spend reduction available anywhere in this doc. If no, gpt-6.1-sol is a roster row and nothing more, and this doc should stop presenting its price as an opportunity. Gates before any volume routing regardless: an independent coding number (every capability claim is OpenAI's harness), an MRCR re-test (the gpt-5.6-luna 41.3 collapse has never been retested anywhere in the gpt-6 line), and the vendor announcement actually fetched (openai.com 403'd this run).
- Is the OpenAI seat worth retargeting every week? (new 2026-10-02, a process question not a model question): the council OpenAI seat has now been proposed for retarget three times in six weeks β gpt-5.5 β gpt-5.6-sol (08-13, never applied) β gpt-6-sol (09-23, applied 09-29) β gpt-6.1-sol (today, seven days later). Each individual swap is correct, same-price, and one line. But the pattern says OpenAI is shipping faster than this flow's cadence, and a seat that is always one release behind is not obviously worse than a seat that churns. Options: keep retargeting (status quo, one line a week), pin the seat to a tier rather than an id and resolve it at call time, or accept a documented lag and only move on a real capability gap. bio-zack's call β I lean toward keep-retargeting while the swaps stay free, but it should be a decision rather than a default.
- The audit roster has now gone unfixed across two reviews (new 2026-10-02): rec 5 on 09-29 was "a data edit to two structures, not a rewrite" and it did not happen; the same gauge lied the same way today. Two readings, and I cannot distinguish them from here: either the fix is lower-priority than the review's framing implied (fine β then this doc should stop calling
aloop_stale=[]a finding and start calling it a known-blind instrument), or the audit's roster needs to stop being hand-maintained at all and readCURRENT_ROSTERout of this doc's roster table, which is the thing that actually gets updated every run. The second is more work once and zero work forever, and it would have caught both the sonnet-5 misclassification and the gpt-6.1-sol miss. Worth a decision rather than a third re-assertion. - grok-4.20 honesty seat β FIRST REAL CHALLENGE (new 2026-09-23, bio-zack's call): grok-4.20 costs $20/$60, the most expensive entry on the roster, and its seat has exactly one justification: an AA-Omniscience honesty lead (recorded here as 78%). Since 2026-04 this doc has declined grok-4.5, 4.6, 4.7, kimi-k3, Fable 5.1, muse-spark-1.3 and gpt-6-astra for a council seat on the same test β "no published AA-Omniscience component." Claude Opus 5.5 is the first model to clear that gate: AA lists it as leading AA-Omniscience outright. Do NOT read that as an automatic retirement, for three reasons. (1) The lead is a different KIND of lead. AA's own benchmark write-up frames Anthropic's Omniscience standing as calibration-driven β low hallucination at moderate accuracy β while xAI's is accuracy-driven β highest raw % correct with a high hallucination rate that the index penalises. A council wants the accuracy pole represented even if it scores lower on the composite. (2) Vendor diversity is itself the point of a fan-out; collapsing the honesty seat into the Claude seat makes it a 3-seat council with two Anthropic votes. (3) The numbers are not on the same footing yet β the 78% figure is old, and the isolated Opus-5.5 Omniscience component was not broken out on AA's model page this run (only "leads" in the launch article). Before deciding, pull isolated AA-Omniscience Index + accuracy + hallucination-rate figures for opus-5-5, grok-4.20 and grok-4.7 on the current scale. If Opus 5.5 leads on BOTH calibration and accuracy, the seat is genuinely redundant and retiring it saves ~$15-20/mo on council fan-out; if it only leads on calibration, keep grok-4.20 and note that the seat's rationale has narrowed. Either way this supersedes the long-stale "Grok 4.3 vs 4.20 retirement decision (~2026-05-14)" bullet below, which never resolved because the 4.3 Omniscience component was never published.
- opus-5-5 spec/behaviour confirmations pending (new 2026-09-23): (a) the anti-distillation safeguard β Anthropic ships a "preserved thinking" anti-distillation measure on API accounts created after 2026-08-31; vita's account predates that by years, so it probably does not apply, but confirm before assuming opus-5 thinking-block behaviour carries over (this bites any code that reads or logs thinking blocks). (b)
mediumis the default effort where opus-5 callers may have relied on a different default β audit any opus-5 caller that never setseffortbefore swapping. (c) thinking cannot be disabled β thethinking:{type:disabled}workaround the 09-21 haikuβsonnet-5 swap needed atfeed/constitutions/extractor.pyis not available on opus-5-5, so never route a small-max_tokensstructured-JSON path here. (d) No independent (non-AA, non-vendor) SWE-bench Pro/Verified reproduction exists yet β the agentic-coding 10 rests on the AA composite. (e) No independent OCR/document-VQA number; vision held at 9. (f) AA has it BEHIND on AA-LCR, its long-context-reasoning component, despite a confirmed 1M window β do not treat 1M nominal as depth. - gpt-6-sol / gpt-6-luna follow-ups (new 2026-09-23): neither was fetched at an OpenAI model card this run (Sol via AA's release page + coverage, Luna via llm-stats/OpenRouter snippets). Pull for Sol: an independent coding number, an isolated AA-Omniscience read (relevant if the OpenAI seat swap lands), and confirmation that the METR reward-hacking finding raised against gpt-5.6-sol does or does not carry β that flag has gated the OpenAI seat swap since 07-10 and it has never been checked at primary (metr.org 403'd on the 07-10 attempt). Pull for Luna: an AA Intelligence Index and an MRCR number β gpt-5.6-luna collapsed at 41.3 despite a nominal 1M window, and until that is re-tested Luna is disqualified from every long-context role, which is most of the budget band's work. Also open: at $0.10/$0.50 Luna undercuts gemini-3.1-flash-lite on output, so if the MRCR re-test reads clean it is the first credible challenger to the budget-fast pin in months.
- deepseek-v4.1-flash peak/off-peak modelling (new 2026-09-23): the first roster entry whose price depends on the hour β off-peak $0.15/$0.60, double at peak, peak MonβFri 01:00β04:00 + 06:00β10:00 UTC. vita's highest-volume scheduled work sits inside the second window (nightly rollup 00:05 PT = 07:05 UTC; /sleep 00:30 PT = 07:30 UTC), so a naive swap from the incumbent v4-flash ($0.14/$0.28 flat) would cost more on exactly the jobs that drive the volume. Before any routing: confirm the schedule at the model card (all pricing is snippet-sourced today), then compute actual spend from
data/inference/usage.jsonltimestamps against both bands. It may still win on the off-peak-heavy modes β but that is a measurement, not an assumption. - claude-opus-5-5 rumor watch β RESOLVED 2026-09-23, and the 09-21 call was WRONG: the model shipped 09-22. The queued re-check (
proactive.queueid 12657) fired as designed and found it at every primary source, with the "leaked" price sheet exact on all four numbers. Two things to carry forward. (1) The discovery-gate tightening proposed on 09-21 β "require a platform.claude.com or openrouter hit beforenew_models_detectedfires" β should NOT be implemented as written. It would have suppressed this detection entirely: the rumor preceded the docs by a day, and the flow's own research step is what verified it. The gate belongs on assertion, not on detection β detect freely, quarantine in the doc, never write "does not exist" from a negative search. (2) This is the second time a dismissed model-status rumor proved real (Fable-5 export-control suspension: dismissed 06-23, confirmed 07-05). One instance is noise; two is a pattern, and the pattern is that this review's failure mode is over-confident negative claims, not over-credulous positive ones. The 09-21 review did the hard part right (it withheld the roster row for lack of a primary source) and then spent a paragraph asserting non-existence it could not support. Withholding needs no such justification β say "unverified, not rostered, re-check" and stop. - ~~claude-opus-5-5 rumor watch (new 2026-09-21):~~ (superseded by the bullet above β kept for the reasoning trail) the 9-21 trigger was a single-source X rumor (codename "claude-wafer-eap", claimed release Tue 9-22, leaked $4/$20 sheet = exactly 20% off Opus 5, demo retracted as recycled footage). Verified NON-EXISTENT at every primary source this run. Re-check 9-23. If it ships: re-run this flow from the announcement + model card; the leaked sheet is not a source. If it doesn't: the discovery-ingest path should learn that a "codename + Tuesday" post with no docs is not a new-model detection β consider requiring a platform.claude.com or openrouter hit before
new_models_detectedfires (the Fable-5 export-control lesson cuts both ways: verify rumors, but don't let them drive the flow). UPDATE 2026-09-21 (evening): the 9-23 re-check is no longer a note-to-self β it is queued indata/proactive/queue.dbas id 12657, triggermodel_guidance_recheck, firing 9-23 9:00am PT with the primary-source-only instruction and both branches (still absent β tighten the discovery gate silently; shipped β re-run the flow from the real announcement and tell him). - Sonnet 4.5 retirement β RESOLVED 2026-10-02, and the four-review prediction was right. A formal deprecation notice landed 2026-09-30 (one day after the floor passed, one day after the last refresh) with retirement 2026-11-30 and
claude-sonnet-5-5as the replacement. Four consecutive reviews carried the floor and said the 60-day rule pushed any real failure to ~late November; it did, almost to the day. Exposure was nil β no live routing pins,geo.pyalready on sonnet-5, only historical price rows remain. The transferable lesson, now with evidence: the notice arrives shortly AFTER the floor passes, not before. Soclaude-haiku-4-5's 10-15 floor should be read as "expect a notice in mid-to-late October, with retirement ~mid-December" rather than "nothing is happening". Zero live haiku pins, so this is a watch, not a risk. - Sonnet 4.5 / Haiku 4.5 retirement floors (new 2026-09-21):
claude-sonnet-4-5-20250929"not sooner than 2026-09-29",claude-haiku-4-5-20251001"not sooner than 2026-10-15" β the first near-term floors on any model vita still pins. No deprecation notice yet, and Anthropic's 60-day rule means nothing fails before ~late Nov, but a notice can arrive any day after the floor. Live pins: sonnet-4-5 atscripts/seo/geo.py:41(β sonnet-5, the current Free/Pro default, which is what the GEO panel is meant to emulate); haiku-4-5 atscripts/feed/constitutions/extractor.py:34,scripts/media/bookmarks.py:148,flows/research-proposal/FLOW.yaml:41. No Haiku 5 exists β the successor decision is sonnet-5 ($2/$10, 2Γ haiku's $1/$5) vs gemini-3.1-flash-lite (~$0.10/$0.40, different vendor) per site. Decide before a notice lands, not after. - Sonnet 5 pilot re-decision (updated 2026-09-21): the 07-05 design said "re-decide at standard pricing (Sep 1) since the intro discount evaporates then" β it didn't evaporate; $2/$10 is permanent and the official tokenizer inflation is ~30%, so sonnet-5 is cost-NEUTRAL vs 4.6 rather than cost-worse. That removes the cost objection; the voice-drift gate is the only one left. Pilot target unchanged:
media_deep_dive(+epd47_generate) in.aloop/config.json, both non-voice-critical, both currently sonnet-4-6. Voice paths (coaching, podcast, silicon-zack chat) stay on 4.6 until a drift check reads clean. - grok-4.7 unknowns (new 2026-09-21): the announcement omits context window, knowledge cutoff, cached-input price, modality, and whether grok-4.6's >200K-token whole-request rebill ($4/$12) carries. AA lists 500K ctx. Confirm all five at the model card / OpenRouter model page before any long-prompt or vision routing. Also reconcile AA's $3.74/task (v4.3.2) vs grok-4.6's $0.84 (v4.1.1) β different index versions, do not read as a 4Γ regression without a same-scale number.
- AA v4.3.2 full-grid rebuild (new 2026-09-21): the 1-10 grid's general-reasoning column and every mini-block now mix FOUR index scales (old / v4.1 / v4.1.1 / v4.3.2). Next non-trigger refresh should re-pull v4.3.2 for every rostered model in one pass and re-score the column once, tagging the scale in the header. Until then, treat relative order within a single scale as valid and cross-scale deltas as noise.
- ~~Should we add Qwen 3.5 to the roster?~~ RESOLVED / superseded 2026-08-05: the Qwen line landed on the roster via Qwen3.8-Max (GA 08-03) β see the new open questions below. (Qwen 3.5 itself was never evaluated and is now moot; 3.8-Max is the current flagship.)
- qwen3.8-max adoption gates (new 2026-08-05, WATCH): rostered but NOT routed anywhere. Three gates before ANY pilot: (1) at least one INDEPENDENT coding number (SWE-bench Pro/Verified or a common-harness Terminal-Bench) β every coding number today is Alibaba's own table; (2) open weights + a named license land (~week of Aug-10) AND a second OpenRouter provider stabilizes capacity (single-provider at launch, K3-style risk); (3) a directly-confirmed AA Intelligence Index (the 53 is a search snippet β the AA model page 404'd on direct fetch) plus an AA vision/OCR number for the multimodal lane. Re-run this review at the weights drop (the /sleep new-model watch should catch coverage; else manual
stepwise run flows/model-guidance-review). Do NOT adopt on vendor benchmarks + a search-snippet index. - qwen3.8-max cheap-multimodal pilot candidacy (new 2026-08-05, bio-zack's call, gated): the ONE lane where 3.8-Max is differentiated is vision β #2 Vision Arena + native text/image/VIDEO at $2/$6 (vs gemini-3.1-pro's $2/$12). IF an AA/independent document-VQA or OCR number confirms the vision quality AND weights land (so health/CGM/labs paths can self-host past the China data-jurisdiction risk), the natural bounded pilot is a NON-sensitive vision path (receipt/screenshot extraction that isn't health data) as a gemini-3.1-pro cost challenger. Not now β Arena rank is not an OCR benchmark, and the sensitive-data + jurisdiction constraints rule out the health vision paths until self-host. Competes with grok-4.5/terra/K3, not the cheap band β do NOT displace k2.6/minimax/glm/flash-lite.
- reflect.md placeholder collision β ELEVATED 2026-08-05 (carried from 07-05, now a live near-miss):
flows/sleep/prompts/reflect.md:342still uses the literalnew_models = 'kimi-k2.7,qwen-3.5'as its example placeholder. This review was triggered by a REAL Qwen model (qwen3.8-max) while the placeholder contains a plausible Qwen name (qwen-3.5) β exactly the spurious-new-model-detection collision the 07-05 note warned about, now one edit-distance from firing. 1-line fix: swap both tokens to obviously-fake strings ('example-model-x'). Higher priority than when first raised (this run is the near-miss). - Cerebras / Groq hosted inference: not yet part of vita's routing. Worth a pass if latency becomes a pain point on silicon-zack-chat.
- DeepSeek V4 post-promo decision (May 6+): at standard $1.74/$3.48, is V4-Pro worth a permanent council seat vs grok-4.20 ($20/$60)? Cheaper, but grok has the unique honesty/Omniscience profile. Decide based on observed contributions during the May-5 promo window.
- V4-Flash bake-off result (~2026-05-07): if quality holds across
podcast-deep/quality_gate.pyandmedia/synthesis.py, fan out to the full 9-mode M2.7 tier. Otherwise document the regression and stay on M2.7. - GPT-5.5-pro policy: currently "bio-zack tag only" β should we automate escalation on council when a question scores above some stakes threshold?
- gpt-5.5-cyber GA watch (new 2026-06-23): evaluated and excluded β gated to verified "trusted defenders" (Daybreak), not on public API/OpenRouter, no published pricing. Strong cyber-domain evals (CyberGym 85.6) but not vita-routable. Re-evaluate ONLY if OpenAI opens general access with pricing. Until then, any autonomous vuln-analysis/patching need falls to the existing coding tier (opus-4-8 / fable-5 / glm-5.2); orgs with vetted access use the GA "GPT-5.5 + Trusted Access for Cyber + Codex Security" path. Don't add a roster/domain-table row (handled like the non-adoptable
claude-mythos-5). - opus-4-7β4-8 bump β DELIBERATE HOLD EXECUTED 2026-07-18 (rule fired; no longer re-surfaced): still unapplied at this review, so per the self-terminating rule below the row is now marked "DELIBERATE HOLD β bio-zack declined by inaction, reaffirm quarterly (~Oct 2026)". The pins stay: aloop
sleep_build,scripts/cadence/daemon.py:258,scripts/reports/generate_html_v2.py:54-55β all opus-4-7, price-neutral to flip whenever he wants. Original context: recommended at 06-17, 06-18, 06-23, 07-05, 07-09, 07-10 and still unapplied β audit-confirmed still live at aloopsleep_build,scripts/cadence/daemon.py:258,scripts/reports/generate_html_v2.py:54-55(the 07-10 audit showsdaemon.py:258= opus-4-7,sleep_build= opus-4-7). Price-neutral ($5/$25 both). The 07-09 review forced a binary; no decision landed. Self-terminating rule, effective now: if still unapplied at the next review, that review auto-marks this row "DELIBERATE HOLD β bio-zack declined by inaction, reaffirm quarterly" and STOPS re-surfacing it. Silence has been the answer five times; the recite loop itself is now the bigger failure mode than the stale pin. One tap to apply, or do nothing and the system stops asking. - grok coding pilot design (new 2026-07-09; retargeted to grok-4.7 2026-09-21, bio-zack approved the 09-21 review): still UNRUN. The pilot arm is now grok-4.7 β 4.5 and 4.6 are both superseded at the same $2/$6 (OpenRouter list price is actually $1.60/$4.80), and the menu refs in
scripts/council/core.py:25+scripts/experiments/model_dialogue.py:29were bumped 4.5 β 4.7 in the same pass. grok's value is cost-per-TASK (not per-token) β a fair bake-off must measure effective cost WITH the token-efficiency applied, not the sticker. AA measures 4.7 at $3.74/task and flags it "notably slow and very verbose", so the bake-off must record wall-clock and output tokens, not just dollars. Candidate bounded path:flows/code-review/synthesize(currently deepseek-v4-pro) as a challenger arm, or a bounded sleep-build sub-agent. Do NOT fan out until one clean side-by-side. Keep OFF council (no Omniscience number) and OFF >500K-context roles (500K ctx). Also unresolved: modality (OpenRouter lists text-only β confirm before any OCR/PDF routing) and whether the SpaceXAI rebrand changed theXAI_API_KEY/base-URL path. grok-4.7 does NOT replace grok-4.3 (4.3 keeps the 1M-ctx xAI reasoning-cheap slot). - Sonnet 5 pilot design (new 2026-07-05): sonnet-4-6 is vita's default chat/composer/voice model, so a wholesale swap is the highest-blast-radius change available β do NOT auto-apply. Recommended: pilot sonnet-5 on ONE non-voice analytical path (
media_deep_dive, which has a quality signal), track effective cost WITH the ~1.0-1.35Γ tokenizer inflation applied (not just per-token rates), and hold all voice-critical + composer paths on 4-6 until the pilot reads clean. Re-decide at standard pricing (Sep 1) since the intro discount evaporates then. Open sub-question: does the tokenizer change alter silicon-zack voice fine-tune assumptions? RESOLVED 2026-09-21 β the pilot is RUNNING.media_deep_dive+epd47_generatemoved to sonnet-5; every voice and composer path stays on sonnet-4-6. The cost question is settled the other way from how this bullet framed it: the $2/$10 intro price is permanent (the 9-01 rise to $3/$15 was cancelled), so with ~30% tokenizer inflation the pilot is roughly cost-flat, not cost-worse β re-deciding "at standard pricing (Sep 1)" is moot. What's left is the voice-drift gate: watch these two modes' output for a week before proposing any wider swap. The open sub-question (does the tokenizer change alter silicon-zack voice fine-tune assumptions?) is untouched and still open. - haiku-4-5 successor β DECIDED 2026-09-21 (bio-zack delegated the call with "make all these recommended changes"): no Haiku 5 exists and the 10-15 retirement floor was approaching with three live pins and no plan. Decided per-site rather than uniformly, because the three sites are not the same kind of work.
scripts/feed/constitutions/extractor.pyβ claude-sonnet-5: it writes constitution amendments behind a 0.7 confidence threshold, so it is identity-adjacent, and staying in-vendor means the prompt shape is unchanged. That swap needed a second line, not one: it calls the Anthropic SDK directly atmax_tokens=400, and sonnet-5 runs adaptive thinking whenthinkingis omitted (haiku-4-5 did not) β thinking would have consumed the budget and truncated the JSON mid-object, so the call now passesthinking={"type": "disabled"}. Any future haikuβ4.6+ swap on a small-max_tokenspath has the same trap.scripts/media/bookmarks.pyβ google/gemini-3.1-flash-lite: already an OpenRouter call, so the vendor switch is genuinely one line, it is 10Γ cheaper, and it matches the tier its neighbours (media_score,media_summarize,rss_process) already use.flows/research-proposal/FLOW.yaml:41β anthropic/claude-sonnet-5: that line is a four-vendor review panel, so flash-lite would have collapsed it onto the Google arm. Note the same line still carriesgpt-4.1-mini,gemini-2.5-flash-preview-05-20anddeepseek-r1β all retired or superseded, none of them in the 09-21 recommendations, and a separate decision. - MiniMax M3 vs the m2.7/v4-flash bake-off (new 2026-07-05): M3 is same price as m2.7 with strictly better specs (1M ctx, native multimodal) β arguably it should just replace m2.7 as the incumbent, making the live bake-off a 3-way (m3 / m2.7 / v4-flash). But benchmarks are unverified and the OpenRouter $0.30/$1.20 may be a launch promo (direct API is $0.60/$2.40 + a 512K-doubling band). Pilot on the
speculative/*m2.7 sites first; verify pricing permanence before any high-volume fan-out. - kimi-k2.7-code has no clean vita swap target (new 2026-07-05): the trigger model is coding-specialized with 256K ctx, but EVERY current kimi-k2.6 site (
docs-audit/review,council-review/synthesize,vdb/generate_newsletter.py,memory/consolidator.py, gumball signal-check) was chosen FOR the 1M window. So k2.7-code can't slot in anywhere today without a context regression. Rostered for future bounded coding agents; k2.6 retained. Re-evaluate if a general (non-Code) K2.7 ships with 1M ctx. (2026-07-18: kimi-k3 IS the general 1M-ctx successor β but at $3/$15 it's a different band entirely; the k2.6 slots stay put. The "cheap 1M Moonshot successor" this bullet waits for still doesn't exist.) - Fable-5 availability fallback (new 2026-07-05): the Jun 12-30 export-control suspension proved the Frontier-max slot can vanish by policy, not just refusal. Before ANY Fable-5 adoption (council-synthesis pilot, sleep_build), the caller must carry an opus-4-8 fallback that triggers on BOTH
stop_reason:refusalAND provider-unavailable/403. Also: mythos-5 remains Glasswing-only. Don't route to Fable-5 without the dual fallback. - ByteDance Seed 2.1 specs pull (new 2026-07-05): confirmed on the llm-stats tracker (GA 2026-06-24, Pro + Turbo variants) but no price/context/benchmark pulled this run β not on vita's roster. Pull specs next refresh and decide whether it's roster-relevant (likely a budget/mid-tier challenger). Do not adopt on zero data.
- Gemini 3.5 Pro (still unconfirmed; GPT-5.6 half RESOLVED 2026-07-10): the 07-05 row held both as unverified rumors β correct call twice over: GPT-5.6 turned out real (GA 07-09, now rostered β though at 1M ctx, not the rumored 1.5M) and the quarantine discipline cost nothing. Gemini 3.5 Pro remains UNCONFIRMED (not checked this run β trigger-scoped research). If real, it's a direct gemini-3.1-pro (Frontier-multimodal seat) bump β verify at next refresh.
- gpt-5.6-sol council-seat swap (new 2026-07-10, bio-zack's call): proposed gpt-5.5 β gpt-5.6-sol on the OpenAI council seat (same $5/$30; sites in the Council section). Gated on bio-zack's OK because of two open flags: (a) METR reward-hacking β launch coverage reports METR measured Sol's detected reward-hacking rate as the highest of any public model it has evaluated; the primary METR post 403'd this run β CONFIRM at metr.org before or shortly after the swap; (b) Omniscience hallucination rate UP vs 5.5 (fine for the reasoning seat while grok-4.20 holds honesty, but worth eyes). gpt-5.5-pro unchanged (no 5.6-pro). Terra note: if the seat swap lands and reads clean, Terra at $2.50/$15 is the natural candidate for any future cost-sensitive frontier-reason path β but one change at a time.
- GPT-Live API watch (new 2026-07-10): gpt-live-1/-mini (full-duplex voice, delegates to frontier text models mid-conversation, replaced Advanced Voice Mode in ChatGPT 07-08) is ChatGPT-only β API "planned", no pricing, no model IDs, not on OpenRouter. NOT rostered (gpt-5.5-cyber/mythos-5 handling: excluded for access, not capability). Vita relevance is real β this is the architecture a silicon-zack voice surface wants β so re-evaluate the moment OpenAI ships API access with pricing. Known launch wrinkles: Hindi translation demo quality issues; Willison's preview hit a laughing-interruption bug; the 75.5 "pleasantness" score is OpenAI-internal.
- Audit CURRENT_ROSTER is frozen β the "0 aloop stale" gauge lies by omission (root-caused 2026-07-10):
flows/model-guidance-review/snapshot_routing.py:27still listsclaude-opus-4-7as CURRENT and lacks opus-4-8, fable-5, glm-5.2, sonnet-5, minimax-m3, kimi-k2.7-code, grok-4.3, grok-4.5, the deepseek v4 family, and now the gpt-5.6 trio. Consequence:aloop_stalerenders empty every review while thesleep_buildβopus-4-7 pin β the exact thing the audit exists to catch β sails through unflagged, anddeep-council/FLOW.yaml:28(opus-4-8, CORRECT) false-positives every run. Four consecutive reviews have footnoted this as a "known blind spot"; it's a 10-minute fix to the set literal. Instrumentation-lying class: a passing check that cannot fail is not a check. - reflect.md placeholder collision (new 2026-07-05):
flows/sleep/prompts/reflect.md:348uses the literal examplenew_models = 'kimi-k2.7,qwen-3.5'as a placeholder β and "kimi-k2.7" is now a real (if -Code-suffixed) model. Low risk, but the placeholder string doubling as a real model name is exactly the kind of thing that can spuriously trigger new-model detection. Consider swapping the placeholder to an obviously-fake token (example-model-x). - Grok 4.3 vs 4.20 retirement decision (~2026-05-14): swap requires (a) AA publishes Omniscience component score β₯75% AND (b) shadow-seat outputs track 4.20 on at least 5 council runs. If both conditions hit, retire 4.20 (saves ~$15-20/mo on council fan-out). If either misses, hold and re-evaluate at next refresh.
- GPT-5.5 reasoning-effort tiers: AA leaderboard now lists
gpt-5.5 (high)= 59 andgpt-5.5 (xhigh)= 60 vs the standardgpt-5.5= 57 currently used in vita. Need to investigate whether OpenAI exposes a request-timereasoning_effortknob and whether council seats should pin to a specific tier. - grok-4.3 context window discrepancy: AA + OpenRouter list 1M; review articles say 2M carryover from 4.20. Treat 1M as the safe operating limit until xAI publishes a model card. Tiered pricing past 200k tokens β verify higher-rate before routing long-context jobs.
- Fable-5 council-synthesis pilot (bio-zack's call): swap the council synthesis model opus-4-8 β claude-fable-5? Highest-leverage adoption (one low-volume call on the highest-stakes step), but requires council code to carry a refusalβopus-4-8 fallback first. Decision gated on that integration + one side-by-side synthesis comparison.
- Fable-5 for
sleep_build: Fable-5 is purpose-built for long-horizon autonomous work (sleep_build's exact profile, VITA #1 persistence surface). Tempting, but 2Γ output cost on a nightly job + adaptive-thinking-always-on + refusal/fallback integration. Recommend: land the conservative opus-4-7β4-8 bump first, observe persistence quality, then consider a Fable-5 pilot. Don't leapfrog. - Fable-5 refusal risk on health/bio surfaces: do NOT route
doctor, health-coaching, or CGM/labs-analysis paths to Fable-5 without fallback β bio-adjacent content can trip its bio safety classifier and returnstop_reason:refusal. This is a sharp, model-specific caveat (Mythos 5 lacks the classifier but is Glasswing-only). Any Fable-5 caller needs the opt-in fallback wired before it sees personal-health content. - Fable-5 benchmark confirmation (PARTIALLY CLOSED 2026-07-10): AA Intelligence Index confirmed direct at artificialanalysis.ai β 60 on Index v4.1, #1 (the "~65" was an old-index single-source figure; scale changed, both were "right" on their own scale). SWE-bench Pro 80 re-confirmed incidentally via the gpt-5.6 comparison coverage. Still open: SWE-bench Verified ~95 remains single-source β low priority now that the Pro number and AA index are both grounded.
- Cost-tracking gap (now 6 reviews old, 2026-07-10 β 10 models missing):
scripts/inference/costs.py(audit-confirmed still only opus-4-5/opus-4-6/sonnet-4-5) +cost_report.pyhave no rows for opus-4-8 ($5/$25), fable-5 ($10/$50), glm-5.2 ($1.40/$4.40), sonnet-5 ($3/$15), minimax-m3 ($0.30/$1.20), kimi-k2.7-code ($0.74/$3.50), grok-4.5 ($2/$6), or now gpt-5.6-sol ($5/$30), gpt-5.6-terra ($2.50/$15), gpt-5.6-luna ($1/$6). New-tier spend currently mis-costs (falls back to nearest-tier estimate). Add all alongside the historical entries (don't replace). Note for the 5.6 rows: cache writes bill at 1.25Γ input β the cost model needs a cache-write column for 5.6+ models or it will under-report. (2026-07-18: add kimi-k3 $3/$15 + cache-hit $0.30 β now 11 models / 7 reviews; Moonshot's cache-hit discount, like the 5.6 cache-write multiplier, needs the cache-aware column.) (2026-09-23 β now ~20 models / 15 reviews, and the cache column is no longer optional.) Add opus-5-5 $4/$20, gpt-6-sol $2/$10, gpt-6-luna $0.10/$0.50. The reason this stopped being pure hygiene: cache-read multipliers have now forked three ways β 0.1Γ standard, 0.05Γ on opus-5-5 ($0.20), 0.025Γ on Fable 5.1/Mythos 5.1 ($0.25). A flat "cache reads are 10% of input" column mis-costs all three, and it mis-costs them in the direction that matters most, because cache-heavy paths are exactly where opus-5-5's price cut actually lands (its per-token cut is cancelled by 1.6Γ verbosity on agentic work β see β). Without a per-model cache multiplier there is no way to measure whether the opus-5β5-5 swap saved anything. Two more shapes the column needs: gpt-6-sol's $0.20 cached read and deepseek-v4.1-flash's time-of-day doubling (if it is ever routed, cost depends on the request's UTC hour, which no current cost model represents). - GLM-5.2 AA Intelligence Index unpublished: SWE-bench Pro 62.1 + the FrontierSWE/MCP-Atlas figures are from Z.ai launch coverage (VentureBeat); artificialanalysis.ai had no GLM-5.2 entry at refresh time and llm-stats listed no benchmark numbers. Re-confirm the coding scores + pull an AA Intel Index next refresh before treating as load-bearing. General-reasoning domain score held at 7 (= glm-5.1) pending AA data.
- GLM-5.2 feed-editor hold revisit:
scripts/feed/narrative.py:144+scripts/feed/privacy.py:170stay on glm-5.1 (cheaper, non-coding structured-editor passes; glm-5.2's gains are coding-specific). Revisit only on a measured quality regression β cheaper-first per methodology. - GLM-5.2 modality: OpenRouter lists glm-5.2 as text-in/out; vision domain score held at 6 (unchanged from 5.1) and it is NOT recommended for any OCR/screenshot/PDF path. Confirm modality before routing image work.
flows/code-review/synthesizedoc-drift: the 2026-04-23 "Sonnet 4.6 swap decisions" table still lists this site as glm-5.1, but it was moved todeepseek/deepseek-v4-proby the 2026-04-30 V4-Pro pilot (flows/code-review/FLOW.yaml:136). Historical table left as-dated; live routing is v4-pro. Not a glm-5.2 bump target.- kimi-k3 weights + license watch (new 2026-07-18): weights PROMISED by 07-27; license unspecified at launch (K2.x was Modified-MIT β don't assume it carries). Watch for: the actual license, independent SWE-bench Pro/Verified or common-harness Terminal-Bench numbers, an AA Omniscience component, OpenRouter capacity stabilizing (frequent 429s at launch), and second providers post-weights (which usually fixes both capacity and price). Re-run this review at the weights drop β the /sleep new-model watch should catch the coverage; if not, manual
stepwise run flows/model-guidance-review. - kimi-k3 AA-rank discrepancy (new 2026-07-18): launch coverage (Bloomberg et al) says K3 debuted #3 on the AA leaderboard, behind only Fable-5 and Sol; AA's own model page says #4 of 187. On the known v4.1 values (Fable-5 60 / Sol 59 / Terra 55) a 57 slots #3 β AA's #4 implies an unlisted β₯57 entry, most plausibly a gpt-5.5 high/xhigh effort-tier row (see the GPT-5.5 reasoning-effort question above β the two puzzles may be one puzzle). Also: Willison's cited "+732 Elo vs K2.6" is implausible as written (fetch artifact); the 1547 Elo itself is indicative (cf. grok-4.5's GDPval-AA 1543). Resolve both at next refresh.
- kimi-k3 frontend-generation pilot (new 2026-07-18, triple-gated): the one lane with an INDEPENDENT K3 #1 (Arena.ai frontend-code, ahead of Fable-5) maps directly onto vita's HTML-report pipeline β candidate sites: the implement stage of
scripts/reports/generate_html_v2.py(currently opus-4.7 under the deliberate hold, ~$3-7/report) or/frontend-slides. Run K3 as a challenger arm on a small report batch, judge with the pipeline's existing critique stage, measure cost-per-report. Gates: weights/license landed + capacity stable + one harness-comparable coding number. Do NOT fold into the grok-4.5/luna SWE-style bounded-coding pilot β different lane, different evidence base. - opus-5 council seat swap (new 2026-07-25, bio-zack's call): bump the Claude deliberation seat + synthesis model opus-4-8 β opus-5. Price-neutral, #1 AA, and it SUPERSEDES the still-open 06-17 Fable-5-synthesis pilot (opus-5 = better/equal synthesis at half Fable-5's cost, ZDR-clean, no export-control availability risk). Only real gate: wire the opus-4-8 cyber-refusal fallback into council code first (health/
doctorcouncils were already OFF Fable-5). Reconcile thecore.py:22opus-4-7 drift in the same edit. Independent of the gpt-5.5βsol OpenAI-seat question β do one seat at a time. - opus-5 vs Fable-5 β the Frontier-max slot may be collapsing (new 2026-07-25): opus-5 is #1 on AA Intelligence Index AND ~26% cheaper-per-task than Fable-5 AND ZDR-clean β so the "hardest builds" default arguably moves from Fable-5 to opus-5. What KEEPS Fable-5 alive: (a) no independent SWE-bench-Pro opus-5-vs-Fable-5 head-to-head was pulled (Fable-5's 80.3 Pro lead over opus-4-8's ~69 is real; whether opus-5 closes it is unknown), (b) Fable-5's verbose deliberate style may still win the rare max-correctness single call. Pull an independent SWE-bench-Pro opus-5 number next refresh; if it matches/beats Fable-5, retire the Frontier-max tier to a footnote. Until then Fable-5 stays as a narrow niche, not a default.
- opus-5 spec confirmations pending (new 2026-07-25): (a) context window β 1M is launch-coverage only; anthropic.com/news + the system-card blurb stated PRICE only this run. Confirm at the model card before any >200K-ctx routing. (b) OpenRouter availability β unverified this run (Anthropic API is day-one). (c) AA Omniscience + TPS/streaming not pulled β the domain mini-block holds honesty at 8 and streaming at 7? pending. (d) opus-5's ZDR-clean status makes it the preferred frontier model for any sensitive/health path where Fable-5's no-ZDR + refusal risk disqualified it β note when routing
doctor/CGM/labs analysis that needs frontier reasoning. - opus-5 for
sleep_build(new 2026-07-25 β HELD, not a re-surface): the held opus-4-7 trio's eventual target is now opus-5 (not opus-4-8), same $5/$25. Per the 07-18 self-terminating rule this is NOT being re-surfaced as an ask β it reaffirms ~Oct 2026, one tap to apply. Recorded here only so the target is correct when the hold next comes up. opus-5 (agentic-coding + computer-use leader, ZDR-clean) is a materially better sleep_build fit than either opus-4-7 or a Fable-5 pilot.