@molecule/api-resource-ai-models 1.6.3 → 1.8.0

package/README.md

@@ -3,7 +3,7 @@ AUTO-GENERATED — DO NOT EDIT THIS FILE.
 Generated by `mlcl sync-docs` from the package's src/index.ts JSDoc + mlcl/registry.json.
 Edits here are overwritten on the next commit (molecule's pre-commit hook regenerates).
 To change this document, edit the module-level JSDoc in src/index.ts.
-Generated: 2026-09-21T16:34:40.015Z
+Generated: 2026-09-24T07:06:51.686Z
 -->
 
 # @molecule/api-resource-ai-models
@@ -159,6 +159,31 @@ interface ModelDefinition {
    * both together; until then, working-without-reasoning beats 400.
    */
   toolsRequireReasoningOff?: boolean
+  /**
+   * The provider rejects a FORCED tool choice for this model — Anthropic
+   * `tool_choice` `any` / `tool` answer 400 on claude-fable-5-1 and
+   * claude-opus-5-5, and Moonshot rejects `required` while thinking is on,
+   * which it always is on kimi-k2.7-code (all probed live 2026-09-23). `auto`
+   * works and the model still calls the tool. Callers that would force a tool call (Synthase
+   * discovery, starting-point selection) must send `auto` for these models;
+   * molecule-dev resolves that in one place (`toolChoiceParam` in
+   * request-shape.ts) and its live dispatch check probes the forced shape for
+   * every model, so a new model with this restriction fails the check instead
+   * of every discovery turn.
+   */
+  rejectsForcedToolChoice?: boolean
+  /**
+   * The provider rejects a caller-chosen `temperature` for this model —
+   * Anthropic answers 400 "`temperature` is deprecated for this model" on the
+   * Claude 5 family and Opus 4.7/4.8 (Opus 4.6, Sonnet 4.6 and Haiku 4.5 still
+   * accept it), and Moonshot's native host allows only 1 on kimi-k3 and
+   * kimi-k2.7-code (all probed live 2026-09-23). Omitting it is always safe. Callers that set a temperature (commit
+   * messages, starting-point selection) must omit it for these models;
+   * molecule-dev resolves that in one place (`temperatureParam` in
+   * request-shape.ts), and its live dispatch check sends the tool-less,
+   * temperature-0 shape to every model.
+   */
+  rejectsTemperature?: boolean
   /**
    * Provider-specific server tool type for web search (e.g. `'web_search_20250305'`).
    * When set, the chat handler sends this as a ServerTool alongside custom tools.
@@ -265,10 +290,26 @@ interface ModelDefinition {
    * over-bill this field exists to prevent. A wrapping window belongs to the
    * day it STARTS on, so its post-midnight tail is still matched against the
    * previous day.
+   *
+   * `excludedDatesUtc` lists `YYYY-MM-DD` dates on which no window applies —
+   * a provider's public holidays. DeepSeek's peak excludes Chinese public
+   * holidays (its pricing page, 2026-09-23); without the list those days bill
+   * at 2× while the provider charges off-peak. A date is matched against the
+   * day a window STARTS on (UTC), like `daysOfWeekUtc`. The list is DATA that
+   * runs out: check-model-freshness warns when the coming year has no dates.
+   *
+   * `rule` is the provider's own sentence defining these windows, verbatim,
+   * and the page that publishes it. check-model-freshness re-reads the page on
+   * every run and warns the moment the sentence changes — the windows are only
+   * as right as the last reading, and DeepSeek has amended its rule twice
+   * (a weekday qualifier by 2026-08-31, a holiday exclusion by 2026-09-23)
+   * without anything here noticing.
    */
   peakPricing?: {
     windows: { startMinuteUtc: number; endMinuteUtc: number; daysOfWeekUtc?: number[] }[]
     multiplier: number
+    excludedDatesUtc?: string[]
+    rule?: { url: string; text: string }
   }
   /**
    * A price change the provider has ANNOUNCED with a dated effective instant,
@@ -317,6 +358,8 @@ interface ModelDefinition {
     peakPricing?: {
       windows: { startMinuteUtc: number; endMinuteUtc: number; daysOfWeekUtc?: number[] }[]
       multiplier: number
+      excludedDatesUtc?: string[]
+      rule?: { url: string; text: string }
     }
     /** Where the change was announced, for the re-verify pass after it lands. */
     source?: string
@@ -543,6 +586,8 @@ function effectivePeakPricing(
   | {
       windows: { startMinuteUtc: number; endMinuteUtc: number; daysOfWeekUtc?: number[] }[]
       multiplier: number
+      excludedDatesUtc?: string[]
+      rule?: { url: string; text: string }
     }
   | undefined
 ```
@@ -786,7 +831,8 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
 `npm run check:model-freshness` from the workspace root):
 
 - Anthropic: https://platform.claude.com/docs/en/about-claude/models/overview
-  - /docs/en/build-with-claude/effort (fable-5 / opus-5 / sonnet-5 current;
+  - /docs/en/build-with-claude/effort (fable-5-1 / opus-5-5 / sonnet-5
+    current as of 2026-09-23 — see the dated notes below; historically fable-5 / opus-5 / sonnet-5 current;
     opus-4-8 superseded by opus-5 at identical pricing but still served — it is
     the recommended refusal-fallback model; effort ladder on all three current
     models is low|medium|high|xhigh|max; budget_tokens 400s on 4.7+)
@@ -816,6 +862,17 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
     any/tool now 400s, thinking blocks are model-bound, editing earlier turns
     invalidates them) — Synthase sends tool_choice auto and is append-only,
     and the pre-commit dispatch probe covers the entry as sent.)
+    (verified 2026-09-23 — ADDED claude-opus-5-5, released 2026-09-22 and
+    listed as "Active (latest)" on /docs/en/models/opus-5-5/overview; the
+    overview page now lists claude-opus-5 under "Legacy models (still
+    available)" → superseded, with opus-4-8/4-7/4-6 repointed one hop.
+    Pricing page: $4/$20, 5m cache write $5, cache hits $0.20 (footnote 2:
+    0.05× input on Opus 5.5 only), fast mode $8/$40. 1M ctx / 128K out, text +
+    image input, reliable knowledge cutoff Jun 2026. Effort page: all five
+    levels, default MEDIUM. What's-new page: thinking always on
+    (disabled/budget_tokens 400), tool_choice any/tool 400, computer_20251124
+    400 — Synthase sends none of those. Every other Anthropic price on the
+    pricing page is unchanged.)
 - OpenAI: https://developers.openai.com/api/docs/pricing (GPT-5.6 family GA
   2026-07-09; REPRICED 2026-07-30: -luna cut 80% to $0.20/$1.20, -terra cut
   20% to $2/$12, -sol unchanged $5/$30; cache read 0.1× input; gpt-5.5/
@@ -836,10 +893,20 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
   request ($20/$75, cached $2, cache writes $25): 2× input/cache, 1.5×
   output. The catalog's price fields are flat per-MTok rates with no
   context-band dimension, so that band is NOT modeled, exactly as the
-  > 200K tiers above are not. It is moot for now — astra is deliberately NOT
-  > in the catalog because it cannot serve a tool-carrying request on the
-  > bond's /v1/chat/completions endpoint; see the DO NOT ADD block above the
-  > OpenAI entries for the live 400s and the condition that lifts it.)
+  > 200K tiers above are not. Astra was withheld at the time: it could not serve a
+  > tool-carrying request on /v1/chat/completions.)
+  > (re-verified 2026-09-24 on the gpt-6-astra model page: ADDED gpt-6-astra
+  > ($10/$50, cached $1, cache writes $12.50; 1.05M ctx / 128K out; image
+  > input; knowledge cutoff Apr 30 2026; effort low|medium|high|xhigh|max),
+  > now that the bond calls /v1/responses.)
+  > (re-verified 2026-09-23: ADDED gpt-6-sol ($2/$10, cached $0.20, cache
+  writes $2.50) and gpt-6-luna ($0.10/$0.50, cached $0.01, cache writes
+  $0.125), released 2026-09-22 — both with the same unmodeled >272K band,
+  > 1.05M ctx / 128K out, image input, knowledge cutoff Apr 20 / May 18 2026
+  > per their model pages. Unlike astra they serve tool-carrying requests on
+  > /v1/chat/completions with reasoning_effort 'none'. The 5.6 family is now
+  > superseded; every 5.6 price on the page is unchanged and the -sol promo
+  > footnote still reads "at least through November 21, 2026".)
 - Google: https://ai.google.dev/gemini-api/docs/pricing (gemini-3.6-flash GA
   2026-07-21 $1.50/$7.50 supersedes 3.5-flash as the agentic flagship;
   gemini-3.1-pro-preview still the pro tier — "3.5 Pro" has NOT shipped as
@@ -872,8 +939,16 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
   not modeled; reasoning_effort low|medium|high default high, image input;
   grok-4.3 still served at $1.25/$2.50 with the bigger 1M window;
   grok-code-fast-1 no longer listed — retires 2026-08-15)
+- Chinese public holidays (DeepSeek's peak excludes them):
+  https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html
+  (the State Council's 2026 notice; see CHINA_PUBLIC_HOLIDAYS)
 - DeepSeek: https://api-docs.deepseek.com/quick_start/pricing +
-  /updates/ (verified 2026-09-10; legacy deepseek-chat/-reasoner ids fully
+  /updates/ (verified 2026-09-23 — the announced 2026-09-14 routing of
+  `deepseek-v4-pro` to V4.1 Flash was WITHDRAWN: V4 Pro keeps its card
+  0.66/1.98/0.022 off-peak, so its staged `scheduledPricing` was deleted.
+  The card now also says peak hours exclude Chinese public holidays, which
+  `peakPricing` cannot express — peak is billed on those weekdays too.
+  Earlier, 2026-09-10: legacy deepseek-chat/-reasoner ids fully
   retired 2026-07-24 — never in this catalog. The 2026-09-10 re-read caught
   the V4.1-Flash release DAY-OF: new evergreen id `deepseek-flash` at
   off-peak miss $0.15 / hit $0.003 / out $0.6 (peak ×2, same Mon-Fri UTC
@@ -958,6 +1033,14 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
   card again — $0.15/$0.50, cached input $0.03, no promo footnote. models.dev
   still carries the expired promo rate, so the KNOWN_DIVERGENCES entry now
   runs to a fixed 2026-12-09 re-verify date.)
+  (verified 2026-09-23: glm-5.3-flashx ADDED — the ~200 tokens/s tier of the
+  GLM-5.3-Flash series. Pricing page: $0.37/$1.25, cached input $0.075,
+  cache storage free. Chat-completion reference: id `glm-5.3-flashx` in the
+  vision request enum, 128K max output, reasoning_effort low|high|max
+  (default max), function tools supported. GLM-5.3-Flash guide: 1M ctx,
+  video/image/text/file input. No published knowledge cutoff. Not on
+  DeepInfra (zai-org/GLM-5.3-FlashX 404s), so cn-region-only. glm-5.3 and
+  glm-5.3-flash rates unchanged on the same page.)
 
 Knowledge-cutoff dates on non-Anthropic entries are best-effort estimates
 where the provider doesn't publish one; the provider sources above verify

package/dist/lookup.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"lookup.d.ts","sourceRoot":"","sources":["../src/lookup.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAGH,OAAO,KAAK,EAAE,YAAY,EAAE,eAAe,EAAE,MAAM,YAAY,CAAA;AAE/D;;;;;;;;GAQG;AACH,wBAAgB,iBAAiB,CAC/B,KAAK,EAAE,IAAI,CAAC,eAAe,EAAE,UAAU,GAAG,cAAc,CAAC,GACxD,OAAO,CAET;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,wBAAwB,CAAC,EAAE,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,CASvE;AAED;;;;;;;GAOG;AACH,eAAO,MAAM,SAAS,EAAE,WAAW,CAAC,MAAM,CAEzC,CAAA;AAED;;;;;;;;;;GAUG;AACH,wBAAgB,QAAQ,CAAC,EAAE,EAAE,MAAM,GAAG,eAAe,GAAG,SAAS,CAEhE;AAED;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAAC,QAAQ,EAAE,YAAY,GAAG,SAAS,eAAe,EAAE,CAEtF;AAED;;;;;;;;;GASG;AACH,wBAAgB,kBAAkB,CAChC,kBAAkB,EAAE,WAAW,CAAC,YAAY,CAAC,GAAG,SAAS,YAAY,EAAE,GACtE,SAAS,eAAe,EAAE,CAI5B;AAoBD;;;;;;;;;;;;GAYG;AACH,wBAAgB,kBAAkB,CAChC,QAAQ,EAAE,eAAe,EACzB,EAAE,GAAE,IAAiB,GACpB,eAAe,CAgBjB;AAED;;;;;;;;;GASG;AACH,wBAAgB,oBAAoB,CAClC,QAAQ,EAAE,eAAe,EACzB,EAAE,GAAE,IAAiB,GACpB,eAAe,CAAC,aAAa,CAAC,CAMhC;AAED;;;;;;;;;;;;;;GAcG;AACH,wBAAgB,oBAAoB,CAClC,QAAQ,EAAE,eAAe,EACzB,EAAE,GAAE,IAAiB,GACpB,eAAe,CASjB;AAED;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,wBAAgB,iBAAiB,CAC/B,QAAQ,EAAE,eAAe,GAAG,SAAS,EACrC,EAAE,EAAE,IAAI,EACR,MAAM,CAAC,EAAE,MAAM,GACd,MAAM,CAsBR;AAED;;;;;;;;;;GAUG;AACH,wBAAgB,oBAAoB,CAClC,QAAQ,EAAE,eAAe,GAAG,SAAS,EACrC,SAAS,CAAC,EAAE,MAAM,GACjB,MAAM,CAGR;AAED,2DAA2D;AAC3D,MAAM,WAAW,eAAe;IAC9B,sDAAsD;IACtD,iBAAiB,EAAE,MAAM,CAAA;IACzB,8CAA8C;IAC9C,kBAAkB,EAAE,MAAM,CAAA;IAC1B,yDAAyD;IACzD,qBAAqB,EAAE,MAAM,CAAA;IAC7B,0DAA0D;IAC1D,sBAAsB,EAAE,MAAM,CAAA;CAC/B;AAED;;;;;;;;;;;;;;;;;;;GAmBG;AACH,wBAAgB,gBAAgB,CAC9B,QAAQ,EAAE,eAAe,EACzB,SAAS,CAAC,EAAE,MAAM,EAClB,EAAE,GAAE,IAAiB,GACpB,eAAe,CAYjB"}
+{"version":3,"file":"lookup.d.ts","sourceRoot":"","sources":["../src/lookup.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAGH,OAAO,KAAK,EAAE,YAAY,EAAE,eAAe,EAAE,MAAM,YAAY,CAAA;AAE/D;;;;;;;;GAQG;AACH,wBAAgB,iBAAiB,CAC/B,KAAK,EAAE,IAAI,CAAC,eAAe,EAAE,UAAU,GAAG,cAAc,CAAC,GACxD,OAAO,CAET;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,wBAAwB,CAAC,EAAE,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,CASvE;AAED;;;;;;;GAOG;AACH,eAAO,MAAM,SAAS,EAAE,WAAW,CAAC,MAAM,CAEzC,CAAA;AAED;;;;;;;;;;GAUG;AACH,wBAAgB,QAAQ,CAAC,EAAE,EAAE,MAAM,GAAG,eAAe,GAAG,SAAS,CAEhE;AAED;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAAC,QAAQ,EAAE,YAAY,GAAG,SAAS,eAAe,EAAE,CAEtF;AAED;;;;;;;;;GASG;AACH,wBAAgB,kBAAkB,CAChC,kBAAkB,EAAE,WAAW,CAAC,YAAY,CAAC,GAAG,SAAS,YAAY,EAAE,GACtE,SAAS,eAAe,EAAE,CAI5B;AAoBD;;;;;;;;;;;;GAYG;AACH,wBAAgB,kBAAkB,CAChC,QAAQ,EAAE,eAAe,EACzB,EAAE,GAAE,IAAiB,GACpB,eAAe,CAgBjB;AAED;;;;;;;;;GASG;AACH,wBAAgB,oBAAoB,CAClC,QAAQ,EAAE,eAAe,EACzB,EAAE,GAAE,IAAiB,GACpB,eAAe,CAAC,aAAa,CAAC,CAMhC;AAED;;;;;;;;;;;;;;GAcG;AACH,wBAAgB,oBAAoB,CAClC,QAAQ,EAAE,eAAe,EACzB,EAAE,GAAE,IAAiB,GACpB,eAAe,CASjB;AAED;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,wBAAgB,iBAAiB,CAC/B,QAAQ,EAAE,eAAe,GAAG,SAAS,EACrC,EAAE,EAAE,IAAI,EACR,MAAM,CAAC,EAAE,MAAM,GACd,MAAM,CA2BR;AAED;;;;;;;;;;GAUG;AACH,wBAAgB,oBAAoB,CAClC,QAAQ,EAAE,eAAe,GAAG,SAAS,EACrC,SAAS,CAAC,EAAE,MAAM,GACjB,MAAM,CAGR;AAED,2DAA2D;AAC3D,MAAM,WAAW,eAAe;IAC9B,sDAAsD;IACtD,iBAAiB,EAAE,MAAM,CAAA;IACzB,8CAA8C;IAC9C,kBAAkB,EAAE,MAAM,CAAA;IAC1B,yDAAyD;IACzD,qBAAqB,EAAE,MAAM,CAAA;IAC7B,0DAA0D;IAC1D,sBAAsB,EAAE,MAAM,CAAA;CAC/B;AAED;;;;;;;;;;;;;;;;;;;GAmBG;AACH,wBAAgB,gBAAgB,CAC9B,QAAQ,EAAE,eAAe,EACzB,SAAS,CAAC,EAAE,MAAM,EAClB,EAAE,GAAE,IAAiB,GACpB,eAAe,CAYjB"}

package/dist/lookup.js

@@ -217,14 +217,20 @@ export function priceMultiplierAt(modelDef, at, region) {
             : minute >= w.startMinuteUtc && minute < w.endMinuteUtc;
         if (!inWindow)
             continue;
+        // A wrapping window belongs to the day it STARTED on, so its
+        // post-midnight tail is matched against the previous UTC day — a
+        // Friday-only 23:00-02:00 window must still be peak at Saturday 00:30.
+        const startedYesterday = wraps && minute < w.endMinuteUtc;
         if (w.daysOfWeekUtc && w.daysOfWeekUtc.length > 0) {
-            // A wrapping window belongs to the day it STARTED on, so its
-            // post-midnight tail is matched against the previous UTC day — a
-            // Friday-only 23:00-02:00 window must still be peak at Saturday 00:30.
-            const day = wraps && minute < w.endMinuteUtc ? (at.getUTCDay() + 6) % 7 : at.getUTCDay();
+            const day = startedYesterday ? (at.getUTCDay() + 6) % 7 : at.getUTCDay();
             if (!w.daysOfWeekUtc.includes(day))
                 continue;
         }
+        if (peak.excludedDatesUtc && peak.excludedDatesUtc.length > 0) {
+            const started = new Date(at.getTime() - (startedYesterday ? 86_400_000 : 0));
+            if (peak.excludedDatesUtc.includes(started.toISOString().slice(0, 10)))
+                continue;
+        }
         return peak.multiplier;
     }
     return 1;

package/dist/models.d.ts

@@ -54,7 +54,8 @@ import type { ModelDefinition } from './types.js';
  * 2026-07-30 GPT-5.6 repricing — cross-check prices against models.dev with
  * `npm run check:model-freshness` from the workspace root):
  * - Anthropic: https://platform.claude.com/docs/en/about-claude/models/overview
- *   + /docs/en/build-with-claude/effort (fable-5 / opus-5 / sonnet-5 current;
+ *   + /docs/en/build-with-claude/effort (fable-5-1 / opus-5-5 / sonnet-5
+ *   current as of 2026-09-23 — see the dated notes below; historically fable-5 / opus-5 / sonnet-5 current;
  *   opus-4-8 superseded by opus-5 at identical pricing but still served — it is
  *   the recommended refusal-fallback model; effort ladder on all three current
  *   models is low|medium|high|xhigh|max; budget_tokens 400s on 4.7+)
@@ -84,6 +85,17 @@ import type { ModelDefinition } from './types.js';
  *   any/tool now 400s, thinking blocks are model-bound, editing earlier turns
  *   invalidates them) — Synthase sends tool_choice auto and is append-only,
  *   and the pre-commit dispatch probe covers the entry as sent.)
+ *   (verified 2026-09-23 — ADDED claude-opus-5-5, released 2026-09-22 and
+ *   listed as "Active (latest)" on /docs/en/models/opus-5-5/overview; the
+ *   overview page now lists claude-opus-5 under "Legacy models (still
+ *   available)" → superseded, with opus-4-8/4-7/4-6 repointed one hop.
+ *   Pricing page: $4/$20, 5m cache write $5, cache hits $0.20 (footnote 2:
+ *   0.05× input on Opus 5.5 only), fast mode $8/$40. 1M ctx / 128K out, text +
+ *   image input, reliable knowledge cutoff Jun 2026. Effort page: all five
+ *   levels, default MEDIUM. What's-new page: thinking always on
+ *   (disabled/budget_tokens 400), tool_choice any/tool 400, computer_20251124
+ *   400 — Synthase sends none of those. Every other Anthropic price on the
+ *   pricing page is unchanged.)
  * - OpenAI: https://developers.openai.com/api/docs/pricing (GPT-5.6 family GA
  *   2026-07-09; REPRICED 2026-07-30: -luna cut 80% to $0.20/$1.20, -terra cut
  *   20% to $2/$12, -sol unchanged $5/$30; cache read 0.1× input; gpt-5.5/
@@ -104,10 +116,20 @@ import type { ModelDefinition } from './types.js';
  *   request ($20/$75, cached $2, cache writes $25): 2× input/cache, 1.5×
  *   output. The catalog's price fields are flat per-MTok rates with no
  *   context-band dimension, so that band is NOT modeled, exactly as the
- *   >200K tiers above are not. It is moot for now — astra is deliberately NOT
- *   in the catalog because it cannot serve a tool-carrying request on the
- *   bond's /v1/chat/completions endpoint; see the DO NOT ADD block above the
- *   OpenAI entries for the live 400s and the condition that lifts it.)
+ *   >200K tiers above are not. Astra was withheld at the time: it could not serve a
+ *   tool-carrying request on /v1/chat/completions.)
+ *   (re-verified 2026-09-24 on the gpt-6-astra model page: ADDED gpt-6-astra
+ *   ($10/$50, cached $1, cache writes $12.50; 1.05M ctx / 128K out; image
+ *   input; knowledge cutoff Apr 30 2026; effort low|medium|high|xhigh|max),
+ *   now that the bond calls /v1/responses.)
+ *   (re-verified 2026-09-23: ADDED gpt-6-sol ($2/$10, cached $0.20, cache
+ *   writes $2.50) and gpt-6-luna ($0.10/$0.50, cached $0.01, cache writes
+ *   $0.125), released 2026-09-22 — both with the same unmodeled >272K band,
+ *   1.05M ctx / 128K out, image input, knowledge cutoff Apr 20 / May 18 2026
+ *   per their model pages. Unlike astra they serve tool-carrying requests on
+ *   /v1/chat/completions with reasoning_effort 'none'. The 5.6 family is now
+ *   superseded; every 5.6 price on the page is unchanged and the -sol promo
+ *   footnote still reads "at least through November 21, 2026".)
  * - Google: https://ai.google.dev/gemini-api/docs/pricing (gemini-3.6-flash GA
  *   2026-07-21 $1.50/$7.50 supersedes 3.5-flash as the agentic flagship;
  *   gemini-3.1-pro-preview still the pro tier — "3.5 Pro" has NOT shipped as
@@ -140,8 +162,16 @@ import type { ModelDefinition } from './types.js';
  *   not modeled; reasoning_effort low|medium|high default high, image input;
  *   grok-4.3 still served at $1.25/$2.50 with the bigger 1M window;
  *   grok-code-fast-1 no longer listed — retires 2026-08-15)
+ * - Chinese public holidays (DeepSeek's peak excludes them):
+ *   https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html
+ *   (the State Council's 2026 notice; see CHINA_PUBLIC_HOLIDAYS)
  * - DeepSeek: https://api-docs.deepseek.com/quick_start/pricing +
- *   /updates/ (verified 2026-09-10; legacy deepseek-chat/-reasoner ids fully
+ *   /updates/ (verified 2026-09-23 — the announced 2026-09-14 routing of
+ *   `deepseek-v4-pro` to V4.1 Flash was WITHDRAWN: V4 Pro keeps its card
+ *   0.66/1.98/0.022 off-peak, so its staged `scheduledPricing` was deleted.
+ *   The card now also says peak hours exclude Chinese public holidays, which
+ *   `peakPricing` cannot express — peak is billed on those weekdays too.
+ *   Earlier, 2026-09-10: legacy deepseek-chat/-reasoner ids fully
  *   retired 2026-07-24 — never in this catalog. The 2026-09-10 re-read caught
  *   the V4.1-Flash release DAY-OF: new evergreen id `deepseek-flash` at
  *   off-peak miss $0.15 / hit $0.003 / out $0.6 (peak ×2, same Mon-Fri UTC
@@ -226,6 +256,14 @@ import type { ModelDefinition } from './types.js';
  *   card again — $0.15/$0.50, cached input $0.03, no promo footnote. models.dev
  *   still carries the expired promo rate, so the KNOWN_DIVERGENCES entry now
  *   runs to a fixed 2026-12-09 re-verify date.)
+ *   (verified 2026-09-23: glm-5.3-flashx ADDED — the ~200 tokens/s tier of the
+ *   GLM-5.3-Flash series. Pricing page: $0.37/$1.25, cached input $0.075,
+ *   cache storage free. Chat-completion reference: id `glm-5.3-flashx` in the
+ *   vision request enum, 128K max output, reasoning_effort low|high|max
+ *   (default max), function tools supported. GLM-5.3-Flash guide: 1M ctx,
+ *   video/image/text/file input. No published knowledge cutoff. Not on
+ *   DeepInfra (zai-org/GLM-5.3-FlashX 404s), so cn-region-only. glm-5.3 and
+ *   glm-5.3-flash rates unchanged on the same page.)
  *
  * Knowledge-cutoff dates on non-Anthropic entries are best-effort estimates
  * where the provider doesn't publish one; the provider sources above verify

package/dist/models.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"models.d.ts","sourceRoot":"","sources":["../src/models.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,YAAY,CAAA;AAQjD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8NG;AACH,eAAO,MAAM,MAAM,EAAE,SAAS,eAAe,EAg3DnC,CAAA"}
+{"version":3,"file":"models.d.ts","sourceRoot":"","sources":["../src/models.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,YAAY,CAAA;AAwDjD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoQG;AACH,eAAO,MAAM,MAAM,EAAE,SAAS,eAAe,EAqkEnC,CAAA"}

package/dist/models.js

@@ -12,6 +12,51 @@
  * prices by business day rather than by hour alone.
  */
 const WEEKDAYS_UTC = [1, 2, 3, 4, 5];
+/**
+ * DeepSeek's own sentence defining its peak windows, verbatim (pricing page,
+ * re-read 2026-09-23). check-model-freshness warns when the page stops saying
+ * exactly this — then the windows and CHINA_PUBLIC_HOLIDAYS need re-reading.
+ */
+const DEEPSEEK_PEAK_RULE = {
+    url: 'https://api-docs.deepseek.com/quick_start/pricing',
+    text: 'Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday, excluding Chinese public holidays.',
+};
+/**
+ * Every date from `from` to `to` inclusive, as YYYY-MM-DD.
+ *
+ * @param from - First date (YYYY-MM-DD).
+ * @param to - Last date (YYYY-MM-DD).
+ * @returns The dates.
+ */
+function dateRange(from, to) {
+    const out = [];
+    for (let t = Date.parse(`${from}T00:00:00Z`); t <= Date.parse(`${to}T00:00:00Z`); t += 86_400_000)
+        out.push(new Date(t).toISOString().slice(0, 10));
+    return out;
+}
+/**
+ * Chinese public holidays (the 放假 ranges), for `peakPricing.excludedDatesUtc`
+ * on the DeepSeek models: their peak is "01:00 - 04:00 and 06:00 - 10:00 UTC,
+ * Monday through Friday, excluding Chinese public holidays" (pricing page,
+ * 2026-09-23). Those windows are 09:00–12:00 and 14:00–18:00 Beijing time, so
+ * the Beijing holiday date IS the UTC date of every window.
+ *
+ * Source — the State Council's notice for 2026 (国办发明电〔2025〕7号), read
+ * 2026-09-23 at https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html.
+ * Make-up working days (调休 weekends) need no entry: the peak is Mon–Fri only.
+ *
+ * ADD NEXT YEAR'S DATES when the State Council publishes them (each November);
+ * check-model-freshness warns 45 days before a year with no dates here.
+ */
+const CHINA_PUBLIC_HOLIDAYS = [
+    ...dateRange('2026-01-01', '2026-01-03'), // 元旦
+    ...dateRange('2026-02-15', '2026-02-23'), // 春节
+    ...dateRange('2026-04-04', '2026-04-06'), // 清明节
+    ...dateRange('2026-05-01', '2026-05-05'), // 劳动节
+    ...dateRange('2026-06-19', '2026-06-21'), // 端午节
+    ...dateRange('2026-09-25', '2026-09-27'), // 中秋节
+    ...dateRange('2026-10-01', '2026-10-07'), // 国庆节
+];
 /**
  * All available AI models, grouped by provider, ordered from most to least capable.
  *
@@ -58,7 +103,8 @@ const WEEKDAYS_UTC = [1, 2, 3, 4, 5];
  * 2026-07-30 GPT-5.6 repricing — cross-check prices against models.dev with
  * `npm run check:model-freshness` from the workspace root):
  * - Anthropic: https://platform.claude.com/docs/en/about-claude/models/overview
- *   + /docs/en/build-with-claude/effort (fable-5 / opus-5 / sonnet-5 current;
+ *   + /docs/en/build-with-claude/effort (fable-5-1 / opus-5-5 / sonnet-5
+ *   current as of 2026-09-23 — see the dated notes below; historically fable-5 / opus-5 / sonnet-5 current;
  *   opus-4-8 superseded by opus-5 at identical pricing but still served — it is
  *   the recommended refusal-fallback model; effort ladder on all three current
  *   models is low|medium|high|xhigh|max; budget_tokens 400s on 4.7+)
@@ -88,6 +134,17 @@ const WEEKDAYS_UTC = [1, 2, 3, 4, 5];
  *   any/tool now 400s, thinking blocks are model-bound, editing earlier turns
  *   invalidates them) — Synthase sends tool_choice auto and is append-only,
  *   and the pre-commit dispatch probe covers the entry as sent.)
+ *   (verified 2026-09-23 — ADDED claude-opus-5-5, released 2026-09-22 and
+ *   listed as "Active (latest)" on /docs/en/models/opus-5-5/overview; the
+ *   overview page now lists claude-opus-5 under "Legacy models (still
+ *   available)" → superseded, with opus-4-8/4-7/4-6 repointed one hop.
+ *   Pricing page: $4/$20, 5m cache write $5, cache hits $0.20 (footnote 2:
+ *   0.05× input on Opus 5.5 only), fast mode $8/$40. 1M ctx / 128K out, text +
+ *   image input, reliable knowledge cutoff Jun 2026. Effort page: all five
+ *   levels, default MEDIUM. What's-new page: thinking always on
+ *   (disabled/budget_tokens 400), tool_choice any/tool 400, computer_20251124
+ *   400 — Synthase sends none of those. Every other Anthropic price on the
+ *   pricing page is unchanged.)
  * - OpenAI: https://developers.openai.com/api/docs/pricing (GPT-5.6 family GA
  *   2026-07-09; REPRICED 2026-07-30: -luna cut 80% to $0.20/$1.20, -terra cut
  *   20% to $2/$12, -sol unchanged $5/$30; cache read 0.1× input; gpt-5.5/
@@ -108,10 +165,20 @@ const WEEKDAYS_UTC = [1, 2, 3, 4, 5];
  *   request ($20/$75, cached $2, cache writes $25): 2× input/cache, 1.5×
  *   output. The catalog's price fields are flat per-MTok rates with no
  *   context-band dimension, so that band is NOT modeled, exactly as the
- *   >200K tiers above are not. It is moot for now — astra is deliberately NOT
- *   in the catalog because it cannot serve a tool-carrying request on the
- *   bond's /v1/chat/completions endpoint; see the DO NOT ADD block above the
- *   OpenAI entries for the live 400s and the condition that lifts it.)
+ *   >200K tiers above are not. Astra was withheld at the time: it could not serve a
+ *   tool-carrying request on /v1/chat/completions.)
+ *   (re-verified 2026-09-24 on the gpt-6-astra model page: ADDED gpt-6-astra
+ *   ($10/$50, cached $1, cache writes $12.50; 1.05M ctx / 128K out; image
+ *   input; knowledge cutoff Apr 30 2026; effort low|medium|high|xhigh|max),
+ *   now that the bond calls /v1/responses.)
+ *   (re-verified 2026-09-23: ADDED gpt-6-sol ($2/$10, cached $0.20, cache
+ *   writes $2.50) and gpt-6-luna ($0.10/$0.50, cached $0.01, cache writes
+ *   $0.125), released 2026-09-22 — both with the same unmodeled >272K band,
+ *   1.05M ctx / 128K out, image input, knowledge cutoff Apr 20 / May 18 2026
+ *   per their model pages. Unlike astra they serve tool-carrying requests on
+ *   /v1/chat/completions with reasoning_effort 'none'. The 5.6 family is now
+ *   superseded; every 5.6 price on the page is unchanged and the -sol promo
+ *   footnote still reads "at least through November 21, 2026".)
  * - Google: https://ai.google.dev/gemini-api/docs/pricing (gemini-3.6-flash GA
  *   2026-07-21 $1.50/$7.50 supersedes 3.5-flash as the agentic flagship;
  *   gemini-3.1-pro-preview still the pro tier — "3.5 Pro" has NOT shipped as
@@ -144,8 +211,16 @@ const WEEKDAYS_UTC = [1, 2, 3, 4, 5];
  *   not modeled; reasoning_effort low|medium|high default high, image input;
  *   grok-4.3 still served at $1.25/$2.50 with the bigger 1M window;
  *   grok-code-fast-1 no longer listed — retires 2026-08-15)
+ * - Chinese public holidays (DeepSeek's peak excludes them):
+ *   https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html
+ *   (the State Council's 2026 notice; see CHINA_PUBLIC_HOLIDAYS)
  * - DeepSeek: https://api-docs.deepseek.com/quick_start/pricing +
- *   /updates/ (verified 2026-09-10; legacy deepseek-chat/-reasoner ids fully
+ *   /updates/ (verified 2026-09-23 — the announced 2026-09-14 routing of
+ *   `deepseek-v4-pro` to V4.1 Flash was WITHDRAWN: V4 Pro keeps its card
+ *   0.66/1.98/0.022 off-peak, so its staged `scheduledPricing` was deleted.
+ *   The card now also says peak hours exclude Chinese public holidays, which
+ *   `peakPricing` cannot express — peak is billed on those weekdays too.
+ *   Earlier, 2026-09-10: legacy deepseek-chat/-reasoner ids fully
  *   retired 2026-07-24 — never in this catalog. The 2026-09-10 re-read caught
  *   the V4.1-Flash release DAY-OF: new evergreen id `deepseek-flash` at
  *   off-peak miss $0.15 / hit $0.003 / out $0.6 (peak ×2, same Mon-Fri UTC
@@ -230,6 +305,14 @@ const WEEKDAYS_UTC = [1, 2, 3, 4, 5];
  *   card again — $0.15/$0.50, cached input $0.03, no promo footnote. models.dev
  *   still carries the expired promo rate, so the KNOWN_DIVERGENCES entry now
  *   runs to a fixed 2026-12-09 re-verify date.)
+ *   (verified 2026-09-23: glm-5.3-flashx ADDED — the ~200 tokens/s tier of the
+ *   GLM-5.3-Flash series. Pricing page: $0.37/$1.25, cached input $0.075,
+ *   cache storage free. Chat-completion reference: id `glm-5.3-flashx` in the
+ *   vision request enum, 128K max output, reasoning_effort low|high|max
+ *   (default max), function tools supported. GLM-5.3-Flash guide: 1M ctx,
+ *   video/image/text/file input. No published knowledge cutoff. Not on
+ *   DeepInfra (zai-org/GLM-5.3-FlashX 404s), so cn-region-only. glm-5.3 and
+ *   glm-5.3-flash rates unchanged on the same page.)
  *
  * Knowledge-cutoff dates on non-Anthropic entries are best-effort estimates
  * where the provider doesn't publish one; the provider sources above verify
@@ -247,6 +330,9 @@ export const MODELS = [
     {
         id: 'claude-fable-5-1',
         provider: 'anthropic',
+        // temperature → 400 "`temperature` is deprecated for this model" (probed
+        // live 2026-09-23); callers omit it (request-shape.ts temperatureParam).
+        rejectsTemperature: true,
         label: 'Claude Fable 5.1',
         description: 'Most capable Anthropic — frontier reasoning & long-horizon agents',
         contextWindow: 1_000_000,
@@ -260,6 +346,10 @@ export const MODELS = [
         // all five levels are supported (effort docs, verified 2026-09-01).
         // Default/recommended is high; xhigh/max for the most capability-sensitive
         // agentic work, medium/low for routine work.
+        // tool_choice any/tool → 400 "tool_choice: type "tool" and "any" are not
+        // supported for this model" (probed live 2026-09-23). Discovery and
+        // starting-point selection send `auto` for it instead (request-shape.ts).
+        rejectsForcedToolChoice: true,
         supportsVision: true,
         supportsPromptCaching: true,
         supportsTools: true,
@@ -282,6 +372,9 @@ export const MODELS = [
     {
         id: 'claude-fable-5',
         provider: 'anthropic',
+        // temperature → 400 "`temperature` is deprecated for this model" (probed
+        // live 2026-09-23); callers omit it (request-shape.ts temperatureParam).
+        rejectsTemperature: true,
         label: 'Claude Fable 5',
         description: 'Most capable Anthropic — frontier reasoning & long-horizon agents',
         contextWindow: 1_000_000,
@@ -317,9 +410,63 @@ export const MODELS = [
         deprecatedAt: '2026-09-01',
         supersededBy: 'claude-fable-5-1',
     },
+    {
+        id: 'claude-opus-5-5',
+        provider: 'anthropic',
+        // temperature → 400 "`temperature` is deprecated for this model" (probed
+        // live 2026-09-23); callers omit it (request-shape.ts temperatureParam).
+        rejectsTemperature: true,
+        label: 'Claude Opus 5.5',
+        description: 'Anthropic Opus flagship — long-running agentic coding, cheaper than Opus 5',
+        // tool_choice any/tool → 400 "tool_choice: type "tool" and "any" are not
+        // supported for this model" (probed live 2026-09-23). Discovery and
+        // starting-point selection send `auto` for it instead (request-shape.ts).
+        rejectsForcedToolChoice: true,
+        // Fast mode (research preview, Claude API only): $8/$40 per MTok (pricing
+        // page, fast-mode table, verified 2026-09-23). Cache multipliers stack on
+        // the fast input rate: read 0.05× (this model's rate), write 1.25×.
+        fastPricing: {
+            inputPricePerMTok: 8,
+            outputPricePerMTok: 40,
+            cacheReadPricePerMTok: 0.4,
+            cacheWritePricePerMTok: 10,
+        },
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 16_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high', 'xhigh', 'max'],
+        defaultEffortLevel: 'medium',
+        // Adaptive thinking is ALWAYS ON: thinking {type:"disabled"} and manual
+        // budget_tokens both 400 — effort is the only depth control. All five
+        // levels supported; the API default is MEDIUM (not high like opus-5).
+        // tool_choice any/tool 400 on this model (auto/none only), as on
+        // fable-5-1. 512-token prompt-cache minimum. (Model page + what's-new,
+        // verified 2026-09-23.)
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        webSearchToolType: 'web_search_20260209',
+        // Same server-tool versions as the rest of the 4.6+ Anthropic fleet. Not
+        // currently sent by Synthase beyond webSearchToolType.
+        codeExecutionToolType: 'code_execution_20260521',
+        webFetchToolType: 'web_fetch_20260209',
+        inputPricePerMTok: 4,
+        outputPricePerMTok: 20,
+        // Cache READ is a documented exception: 0.05× input ("$0.20 / MTok",
+        // pricing page footnote 2). 5-minute cache write is the usual 1.25×.
+        cacheReadPricePerMTok: 0.2,
+        cacheWritePricePerMTok: 5,
+        // Reliable knowledge cutoff Jun 2026 (model page "Specifications").
+        knowledgeCutoff: '2026-06-01',
+    },
     {
         id: 'claude-opus-5',
         provider: 'anthropic',
+        // temperature → 400 "`temperature` is deprecated for this model" (probed
+        // live 2026-09-23); callers omit it (request-shape.ts temperatureParam).
+        rejectsTemperature: true,
         label: 'Claude Opus 5',
         description: 'Anthropic Opus flagship — step-change agentic coding at 4.8 pricing',
         // Fast mode (research preview, Claude API only): same model at up to 2.5×
@@ -362,10 +509,18 @@ export const MODELS = [
         cacheWritePricePerMTok: 6.25,
         // Not published at verification time — best-effort estimate (≥ Opus 4.8's).
         knowledgeCutoff: '2026-01-01',
+        // Superseded by claude-opus-5-5 (released 2026-09-22, same Opus tier,
+        // cheaper at $4/$20); Anthropic now lists opus-5 under "Legacy models
+        // (still available)". Still served and priceable, just not OFFERED.
+        deprecatedAt: '2026-09-22',
+        supersededBy: 'claude-opus-5-5',
     },
     {
         id: 'claude-opus-4-8',
         provider: 'anthropic',
+        // temperature → 400 "`temperature` is deprecated for this model" (probed
+        // live 2026-09-23); callers omit it (request-shape.ts temperatureParam).
+        rejectsTemperature: true,
         label: 'Claude Opus 4.8',
         description: 'Previous Opus — deep reasoning; the opus-5 refusal fallback',
         contextWindow: 1_000_000,
@@ -396,12 +551,16 @@ export const MODELS = [
         // Superseded by claude-opus-5 (same price, same tier); still served upstream
         // and the recommended refusal-fallback target, so it stays priceable and
         // callable by id — it is just not OFFERED, since opus-5 is a drop-in.
+        // `supersededBy` names the CURRENT selectable Opus (opus-5-5, one hop).
         deprecatedAt: '2026-07-28',
-        supersededBy: 'claude-opus-5',
+        supersededBy: 'claude-opus-5-5',
     },
     {
         id: 'claude-sonnet-5',
         provider: 'anthropic',
+        // temperature → 400 "`temperature` is deprecated for this model" (probed
+        // live 2026-09-23); callers omit it (request-shape.ts temperatureParam).
+        rejectsTemperature: true,
         label: 'Claude Sonnet 5',
         description: 'Fast & capable — near-Opus coding at Sonnet cost',
         contextWindow: 1_000_000,
@@ -437,6 +596,9 @@ export const MODELS = [
     {
         id: 'claude-opus-4-7',
         provider: 'anthropic',
+        // temperature → 400 "`temperature` is deprecated for this model" (probed
+        // live 2026-09-23); callers omit it (request-shape.ts temperatureParam).
+        rejectsTemperature: true,
         label: 'Claude Opus 4.7',
         description: 'Older Opus — long-horizon agentic work, knowledge work & vision',
         contextWindow: 1_000_000,
@@ -470,10 +632,10 @@ export const MODELS = [
         // still Active upstream (deprecations page 2026-08-06: retires no sooner
         // than 2027-04-16). NO fast mode — speed:"fast" on 4.7 returns an error
         // (pricing page, fast-mode section). `supersededBy` names the CURRENT
-        // selectable Opus (opus-5), not the also-superseded 4.8, so a saved
+        // selectable Opus (opus-5-5), not the also-superseded 4.8 / 5, so a saved
         // selection resolves forward in one hop.
         deprecatedAt: '2026-05-28',
-        supersededBy: 'claude-opus-5',
+        supersededBy: 'claude-opus-5-5',
     },
     {
         id: 'claude-opus-4-6',
@@ -505,9 +667,9 @@ export const MODELS = [
         cacheReadPricePerMTok: 0.5,
         cacheWritePricePerMTok: 6.25,
         knowledgeCutoff: '2025-05-01',
-        // Superseded by the current Opus (opus-5) — kept priceable, not offered.
+        // Superseded by the current Opus (opus-5-5) — kept priceable, not offered.
         deprecatedAt: '2026-06-16',
-        supersededBy: 'claude-opus-5',
+        supersededBy: 'claude-opus-5-5',
     },
     {
         id: 'claude-sonnet-4-6',
@@ -586,33 +748,116 @@ export const MODELS = [
     // exist upstream — not modeled (same as the Gemini/Grok >200K tiers), and
     // neither are the Batch (0.5×) or Sol "Fast mode" (2×) cards.
     //
-    // DO NOT ADD gpt-6-astra (OpenAI's flagship since 2026-09-04) UNTIL THE BOND
-    // MOVES TO /v1/responses. It is deliberately absent, not overlooked. The
-    // openai bond posts to /v1/chat/completions, and on that endpoint this model
-    // rejects EVERY request Synthase can send, because Synthase always carries
-    // function tools (probed live 2026-09-21 on our own key):
-    //   - tools + reasoning_effort low|medium|high|xhigh → 400 "Function tools
-    //     with reasoning_effort are not supported for gpt-6-astra in
-    //     /v1/chat/completions. To use function tools, use /v1/responses or set
-    //     reasoning_effort to 'none'."
-    //   - tools, field omitted → the same 400 (the model applies its own default)
-    //   - tools + reasoning_effort 'none' → 400 "Unsupported value: … does not
-    //     support 'none' with this model. Supported values are: 'low', 'medium',
-    //     'high', and 'xhigh'."
-    // So the gpt-5.6 workaround (`toolsRequireReasoningOff`, which pins an
-    // explicit 'none') does NOT carry over: OpenAI removed 'none' from this
-    // model's ladder, which closes the one door that made the 5.6 family usable.
-    // Nothing is wrong with the model or the account — no-tools chat and
-    // /v1/responses WITH tools both return 200. Both candidate entries were built
-    // and run through molecule-dev's verify:model-dispatch; both FAILED, so the
-    // entry was withheld rather than shipped broken (the glm-5.3-flash lesson:
-    // a catalog entry that cannot serve a Synthase-shaped turn breaks every turn
-    // on it). The freshness gate WILL keep listing it as a new-model candidate —
-    // that is correct; it becomes addable the day the bond speaks /v1/responses.
-    // Note also that the docs page advertises effort 'max', which
-    // /v1/chat/completions rejects for this model — verify the ladder against the
-    // endpoint, not the docs, when this is revisited.
+    // gpt-6-astra (OpenAI's flagship since 2026-09-04) was withheld until the
+    // bond moved to /v1/responses: on /v1/chat/completions it rejects every
+    // tool-carrying request (tools + any reasoning_effort → 400, and it has no
+    // 'none' effort to fall back on — probed live 2026-09-21). The bond calls
+    // /v1/responses on OpenAI's own endpoint since 2026-09-24, where tools +
+    // every effort low..max return 200 (probed live 2026-09-24), so astra needs
+    // no `toolsRequireReasoningOff` pin. The docs page lists chat/completions
+    // tool support too; the live endpoint disagrees — trust the endpoint.
+    //
+    // gpt-6-sol and gpt-6-luna (released 2026-09-22) are NOT astra's case: both
+    // still accept reasoning_effort 'none', and tools + 'none' returns 200 on
+    // /v1/chat/completions (probed live 2026-09-23 on our own key), so the same
+    // `toolsRequireReasoningOff` pin that serves the 5.6 family serves them. On
+    // that endpoint both accept none|low|medium|high|xhigh and reject 'max' —
+    // again despite the docs pages listing it. GPT-6 has no Terra tier: sol at
+    // $2/$10 is priced at 5.6-terra's tier and undercuts 5.6-sol, so it supersedes
+    // both; luna supersedes 5.6-luna at half the price.
     // ---------------------------------------------------------------------------
+    {
+        id: 'gpt-6-astra',
+        provider: 'openai',
+        label: 'GPT-6 Astra',
+        description: 'OpenAI flagship — the hardest reasoning & coding work',
+        // Documented as 1.05M; floored to 1M like the other OpenAI entries.
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 16_000,
+        thinkingConfigurable: true,
+        // No 'none' on this model. 'max' verified accepted with tools on
+        // /v1/responses (2026-09-24).
+        supportedEffortLevels: ['low', 'medium', 'high', 'xhigh', 'max'],
+        defaultEffortLevel: 'medium',
+        // temperature → 400 "Unsupported parameter: 'temperature' is not supported
+        // with this model" (probed live 2026-09-24).
+        rejectsTemperature: true,
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // NO webSearchToolType yet: /v1/responses serves `web_search` (verified
+        // 2026-09-24), but its per-call fee is not metered — add with metering.
+        codeExecutionToolType: 'code_interpreter',
+        // Standard tier; >272K band ($20/$75, cached $2, writes $25) not modeled.
+        inputPricePerMTok: 10,
+        outputPricePerMTok: 50,
+        cacheReadPricePerMTok: 1,
+        cacheWritePricePerMTok: 12.5,
+        knowledgeCutoff: '2026-04-30',
+    },
+    {
+        id: 'gpt-6-sol',
+        provider: 'openai',
+        label: 'GPT-6 Sol',
+        description: 'OpenAI for complex coding & agentic work',
+        // Documented as 1.05M; floored to 1M like the 5.6 entries.
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 16_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high', 'xhigh'],
+        defaultEffortLevel: 'medium',
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
+        // The bond now calls /v1/responses, where it is not (verified 2026-09-24);
+        // lifting this pin changes reasoning quality/cost, so it waits for a model eval.
+        toolsRequireReasoningOff: true,
+        // NO webSearchToolType: see gpt-6-astra — served on /v1/responses, but
+        // its per-call fee is not metered yet.
+        codeExecutionToolType: 'code_interpreter',
+        // Standard tier. A long-context band above 272K prompt tokens reprices the
+        // whole request ($4/$15, cached $0.40, cache writes $5) — not modeled, same
+        // as the other >200K tiers.
+        inputPricePerMTok: 2,
+        outputPricePerMTok: 10,
+        cacheReadPricePerMTok: 0.2,
+        cacheWritePricePerMTok: 2.5,
+        knowledgeCutoff: '2026-04-20',
+    },
+    {
+        id: 'gpt-6-luna',
+        provider: 'openai',
+        label: 'GPT-6 Luna',
+        description: 'Fast & cheap OpenAI — light tasks & subagents',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high', 'xhigh'],
+        defaultEffortLevel: 'medium',
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
+        // The bond now calls /v1/responses, where it is not (verified 2026-09-24);
+        // lifting this pin changes reasoning quality/cost, so it waits for a model eval.
+        toolsRequireReasoningOff: true,
+        // NO webSearchToolType: see gpt-5.6-sol.
+        codeExecutionToolType: 'code_interpreter',
+        // Standard tier; >272K band ($0.20/$0.75, cached $0.02, writes $0.25) not
+        // modeled.
+        inputPricePerMTok: 0.1,
+        outputPricePerMTok: 0.5,
+        cacheReadPricePerMTok: 0.01,
+        cacheWritePricePerMTok: 0.125,
+        knowledgeCutoff: '2026-05-18',
+    },
     {
         id: 'gpt-5.6-sol',
         provider: 'openai',
@@ -631,12 +876,11 @@ export const MODELS = [
         supportsPromptCaching: true,
         supportsTools: true,
         // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
+        // The bond now calls /v1/responses, where it is not (verified 2026-09-24);
+        // lifting this pin changes reasoning quality/cost, so it waits for a model eval.
         toolsRequireReasoningOff: true,
-        // NO webSearchToolType: the OpenAI bond calls /v1/chat/completions, which
-        // has no web_search tool type (it is a Responses-API construct), and the
-        // bond deliberately forwards no server tools. Advertising one here surfaced
-        // web search in the system prompt while it could never work. Re-add when
-        // the bond moves to /v1/responses (verified 2026-08-28).
+        // NO webSearchToolType: see gpt-6-astra — served on /v1/responses, but
+        // its per-call fee is not metered yet.
         codeExecutionToolType: 'code_interpreter',
         // LIST price. OpenAI ran a >20% PROMO from 2026-08-22 ($4/$20, cache read
         // $0.40, cache write $5) — "GPT-5.6 Sol's promotional pricing is available
@@ -651,6 +895,10 @@ export const MODELS = [
         cacheWritePricePerMTok: 6.25,
         // Not published — best-effort estimate.
         knowledgeCutoff: '2026-03-01',
+        // Superseded by gpt-6-sol ($2/$10 vs $5/$30 list). Still served and
+        // priceable — just not offered in the picker.
+        deprecatedAt: '2026-09-23',
+        supersededBy: 'gpt-6-sol',
     },
     {
         id: 'gpt-5.6-terra',
@@ -668,12 +916,11 @@ export const MODELS = [
         supportsPromptCaching: true,
         supportsTools: true,
         // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
+        // The bond now calls /v1/responses, where it is not (verified 2026-09-24);
+        // lifting this pin changes reasoning quality/cost, so it waits for a model eval.
         toolsRequireReasoningOff: true,
-        // NO webSearchToolType: the OpenAI bond calls /v1/chat/completions, which
-        // has no web_search tool type (it is a Responses-API construct), and the
-        // bond deliberately forwards no server tools. Advertising one here surfaced
-        // web search in the system prompt while it could never work. Re-add when
-        // the bond moves to /v1/responses (verified 2026-08-28).
+        // NO webSearchToolType: see gpt-6-astra — served on /v1/responses, but
+        // its per-call fee is not metered yet.
         codeExecutionToolType: 'code_interpreter',
         // Repriced 2026-07-30 (20% cut from $2.50/$15).
         inputPricePerMTok: 2,
@@ -683,6 +930,10 @@ export const MODELS = [
         cacheWritePricePerMTok: 2.5,
         // Not published — best-effort estimate.
         knowledgeCutoff: '2026-03-01',
+        // Superseded by gpt-6-sol: GPT-6 has no Terra tier, and sol sits at this
+        // tier's price ($2/$10 vs $2/$12).
+        deprecatedAt: '2026-09-23',
+        supersededBy: 'gpt-6-sol',
     },
     {
         id: 'gpt-5.6-luna',
@@ -700,12 +951,11 @@ export const MODELS = [
         supportsPromptCaching: true,
         supportsTools: true,
         // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
+        // The bond now calls /v1/responses, where it is not (verified 2026-09-24);
+        // lifting this pin changes reasoning quality/cost, so it waits for a model eval.
         toolsRequireReasoningOff: true,
-        // NO webSearchToolType: the OpenAI bond calls /v1/chat/completions, which
-        // has no web_search tool type (it is a Responses-API construct), and the
-        // bond deliberately forwards no server tools. Advertising one here surfaced
-        // web search in the system prompt while it could never work. Re-add when
-        // the bond moves to /v1/responses (verified 2026-08-28).
+        // NO webSearchToolType: see gpt-6-astra — served on /v1/responses, but
+        // its per-call fee is not metered yet.
         codeExecutionToolType: 'code_interpreter',
         // Repriced 2026-07-30 (80% cut from $1/$6).
         inputPricePerMTok: 0.2,
@@ -720,6 +970,9 @@ export const MODELS = [
         // note for the same removal pattern).
         // Not published — best-effort estimate.
         knowledgeCutoff: '2026-03-01',
+        // Superseded by gpt-6-luna ($0.10/$0.50 — half the price).
+        deprecatedAt: '2026-09-23',
+        supersededBy: 'gpt-6-luna',
     },
     {
         id: 'gpt-5.5',
@@ -739,12 +992,11 @@ export const MODELS = [
         supportsPromptCaching: true,
         supportsTools: true,
         // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
+        // The bond now calls /v1/responses, where it is not (verified 2026-09-24);
+        // lifting this pin changes reasoning quality/cost, so it waits for a model eval.
         toolsRequireReasoningOff: true,
-        // NO webSearchToolType: the OpenAI bond calls /v1/chat/completions, which
-        // has no web_search tool type (it is a Responses-API construct), and the
-        // bond deliberately forwards no server tools. Advertising one here surfaced
-        // web search in the system prompt while it could never work. Re-add when
-        // the bond moves to /v1/responses (verified 2026-08-28).
+        // NO webSearchToolType: see gpt-6-astra — served on /v1/responses, but
+        // its per-call fee is not metered yet.
         codeExecutionToolType: 'code_interpreter',
         inputPricePerMTok: 5,
         outputPricePerMTok: 30,
@@ -752,10 +1004,11 @@ export const MODELS = [
         cacheReadPricePerMTok: 0.5,
         cacheWritePricePerMTok: 5,
         knowledgeCutoff: '2025-12-01',
-        // Superseded by gpt-5.6-sol (same frontier tier, same $5/$30); still listed
-        // as current by OpenAI, so it stays priceable — it is just not offered.
+        // Superseded by gpt-5.6-sol (same frontier tier, same $5/$30), and since
+        // 2026-09-23 points one hop to gpt-6-sol, 5.6-sol's own successor. Still
+        // listed as current by OpenAI, so it stays priceable — just not offered.
         deprecatedAt: '2026-07-09',
-        supersededBy: 'gpt-5.6-sol',
+        supersededBy: 'gpt-6-sol',
     },
     {
         id: 'gpt-5.4',
@@ -773,12 +1026,11 @@ export const MODELS = [
         supportsPromptCaching: true,
         supportsTools: true,
         // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
+        // The bond now calls /v1/responses, where it is not (verified 2026-09-24);
+        // lifting this pin changes reasoning quality/cost, so it waits for a model eval.
         toolsRequireReasoningOff: true,
-        // NO webSearchToolType: the OpenAI bond calls /v1/chat/completions, which
-        // has no web_search tool type (it is a Responses-API construct), and the
-        // bond deliberately forwards no server tools. Advertising one here surfaced
-        // web search in the system prompt while it could never work. Re-add when
-        // the bond moves to /v1/responses (verified 2026-08-28).
+        // NO webSearchToolType: see gpt-6-astra — served on /v1/responses, but
+        // its per-call fee is not metered yet.
         codeExecutionToolType: 'code_interpreter',
         inputPricePerMTok: 2.5,
         outputPricePerMTok: 15,
@@ -789,9 +1041,10 @@ export const MODELS = [
         // OpenAI still lists gpt-5.4 as current, but gpt-5.6-terra covers this
         // balanced tier for LESS ($2/$12 vs $2.50/$15) — superseded, so the picker
         // offers only the 5.6 generation (this is OUR taxonomy, not OpenAI's
-        // deprecations page; the model stays priceable).
+        // deprecations page; the model stays priceable). Points one hop to
+        // gpt-6-sol since 2026-09-23, when 5.6-terra was itself superseded.
         deprecatedAt: '2026-07-28',
-        supersededBy: 'gpt-5.6-terra',
+        supersededBy: 'gpt-6-sol',
     },
     {
         id: 'gpt-5.4-mini',
@@ -809,12 +1062,11 @@ export const MODELS = [
         supportsPromptCaching: true,
         supportsTools: true,
         // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
+        // The bond now calls /v1/responses, where it is not (verified 2026-09-24);
+        // lifting this pin changes reasoning quality/cost, so it waits for a model eval.
         toolsRequireReasoningOff: true,
-        // NO webSearchToolType: the OpenAI bond calls /v1/chat/completions, which
-        // has no web_search tool type (it is a Responses-API construct), and the
-        // bond deliberately forwards no server tools. Advertising one here surfaced
-        // web search in the system prompt while it could never work. Re-add when
-        // the bond moves to /v1/responses (verified 2026-08-28).
+        // NO webSearchToolType: see gpt-6-astra — served on /v1/responses, but
+        // its per-call fee is not metered yet.
         codeExecutionToolType: 'code_interpreter',
         inputPricePerMTok: 0.75,
         outputPricePerMTok: 4.5,
@@ -825,9 +1077,10 @@ export const MODELS = [
         // Superseded by gpt-5.6-luna, which IS the newer cheap/fast tier and is
         // strictly better on every axis that made this the budget pick: $0.20/$1.20
         // vs $0.75/$4.50 after the 2026-07-30 repricing, and a 1M window vs 400K.
-        // Hiding it therefore costs OpenAI no cheap option. Stays priceable.
+        // Hiding it therefore costs OpenAI no cheap option. Stays priceable. Points
+        // one hop to gpt-6-luna since 2026-09-23, when 5.6-luna was superseded.
         deprecatedAt: '2026-08-01',
-        supersededBy: 'gpt-5.6-luna',
+        supersededBy: 'gpt-6-luna',
     },
     // ---------------------------------------------------------------------------
     // Google
@@ -1213,8 +1466,10 @@ export const MODELS = [
     //      leans on that "temporary" routing until molecule-dev's default ids
     //      move to `deepseek-flash`.
     //   2. From 2026-09-14 12:00 Beijing (04:00 UTC), until V4.1 Pro ships, ALL
-    //      `deepseek-v4-pro` requests route to V4.1 Flash at Flash prices —
-    //      staged as `scheduledPricing` on the pro entry below.
+    //      `deepseek-v4-pro` requests were to route to V4.1 Flash at Flash
+    //      prices. WITHDRAWN before it landed (re-read 2026-09-23): V4 Pro keeps
+    //      being served after 2026-09-14 "with the billing method remaining
+    //      unchanged", so the staged `scheduledPricing` was deleted.
     // 2026-08-13: V4-Pro GA — and with it the price rise that the "coming soon"
     // note below had been waiting on. It was staged as `scheduledPricing`
     // effective 2026-08-16T16:00Z; that instant has PASSED and the rates are now
@@ -1315,33 +1570,25 @@ export const MODELS = [
                 { startMinuteUtc: 360, endMinuteUtc: 600, daysOfWeekUtc: WEEKDAYS_UTC },
             ],
             multiplier: 2,
+            excludedDatesUtc: CHINA_PUBLIC_HOLIDAYS,
+            rule: DEEPSEEK_PEAK_RULE,
         },
-        // ANNOUNCED 2026-09-10 (updates page): from 2026-09-14 12:00 Beijing time
-        // (04:00 UTC), and until V4.1 Pro ships, every `deepseek-v4-pro` request is
-        // routed to V4.1 Flash and billed at the V4.1 FLASH price — so the base
-        // rates become the flash card below (the peak windows are identical, so
-        // they carry through unchanged). The US `regionPricing` above is DeepInfra's
-        // own card for its V4-Pro copy and is NOT touched by the native routing.
-        // Fold into the base fields once the instant has passed (the freshness
-        // gate gives models.dev its scheduled-landing grace meanwhile).
-        scheduledPricing: {
-            effectiveFrom: '2026-09-14T04:00:00Z',
-            inputPricePerMTok: 0.15,
-            outputPricePerMTok: 0.6,
-            cacheReadPricePerMTok: 0.003,
-            cacheWritePricePerMTok: 0.15,
-            source: 'https://api-docs.deepseek.com/updates/ (2026-09-10): deepseek-v4-pro → V4.1 Flash at Flash prices from 2026-09-14 12:00 Beijing',
-        },
+        // The 2026-09-10 announcement that this id would route to V4.1 Flash at
+        // Flash prices from 2026-09-14 was WITHDRAWN: the updates page now reads
+        // "we have decided to continue providing API services for DeepSeek V4 Pro
+        // after September 14, 2026, with the billing method remaining unchanged",
+        // and the pricing page still lists `deepseek-v4-pro` (V4-Pro-0813) at
+        // 0.66/1.98, cache hit 0.022 off-peak (both re-read 2026-09-23). So the
+        // staged `scheduledPricing` flash card was deleted rather than folded in —
+        // the base rates above are what DeepSeek bills. The US `regionPricing` is
+        // DeepInfra's own card and was never affected.
         // Not published by DeepSeek — best-effort estimate.
         knowledgeCutoff: '2025-07-01',
-        // Deprecated the day it stops being itself: DeepSeek's own testing has
-        // V4.1 Flash "comprehensively surpassing" Pro on performance/cost/speed,
-        // and from 2026-09-14 12:00 Beijing every `deepseek-v4-pro` request routes
-        // to V4.1 Flash at Flash prices (see scheduledPricing above) until V4.1 Pro
-        // ships — at which point this becomes a new entry's succession problem, not
-        // this one's. Until the 14th it still serves real V4-Pro-0813 weights at
-        // the Pro card, so it stays selectable for existing selections.
-        deprecatedAt: '2026-09-14',
+        // Was deprecated 2026-09-14 on the expectation that the id would stop
+        // being itself (routed to V4.1 Flash). DeepSeek withdrew that routing — V4
+        // Pro continues "with the billing method remaining unchanged" (updates
+        // page, re-read 2026-09-23) — so it is offered again (owner, 2026-09-23:
+        // "keep offering deepseek pro if it is available").
     },
     {
         id: 'deepseek-v4-flash',
@@ -1424,6 +1671,8 @@ export const MODELS = [
                 { startMinuteUtc: 360, endMinuteUtc: 600, daysOfWeekUtc: WEEKDAYS_UTC },
             ],
             multiplier: 2,
+            excludedDatesUtc: CHINA_PUBLIC_HOLIDAYS,
+            rule: DEEPSEEK_PEAK_RULE,
         },
         // Not published by DeepSeek — best-effort estimate.
         knowledgeCutoff: '2025-07-01',
@@ -1488,6 +1737,8 @@ export const MODELS = [
                 { startMinuteUtc: 360, endMinuteUtc: 600, daysOfWeekUtc: WEEKDAYS_UTC },
             ],
             multiplier: 2,
+            excludedDatesUtc: CHINA_PUBLIC_HOLIDAYS,
+            rule: DEEPSEEK_PEAK_RULE,
         },
         // Not published by DeepSeek — best-effort estimate (family estimate; the
         // V4.1 announcement lists benchmarks but no training cutoff).
@@ -1514,6 +1765,9 @@ export const MODELS = [
     {
         id: 'kimi-k3',
         provider: 'moonshot',
+        // Native Moonshot host: temperature 0 → 400 "invalid temperature: only 1 is
+        // allowed for this model" (probed 2026-09-23); callers omit it.
+        rejectsTemperature: true,
         label: 'Kimi K3',
         description: 'Moonshot flagship — 2.8T open weights, 1M context, multimodal',
         contextWindow: 1_000_000,
@@ -1557,6 +1811,12 @@ export const MODELS = [
     {
         id: 'kimi-k2.7-code',
         provider: 'moonshot',
+        // Native Moonshot host (probed 2026-09-23): temperature 0 → 400 "only 1 is
+        // allowed"; and tool_choice 'required' → 400 "incompatible with thinking
+        // enabled" — thinking cannot be turned off on this model, so a forced tool
+        // call goes out as `auto` (request-shape.ts).
+        rejectsTemperature: true,
+        rejectsForcedToolChoice: true,
         label: 'Kimi K2.7 Code',
         description: 'Moonshot coding specialist — token-efficient agentic coding',
         contextWindow: 262_144,
@@ -2003,6 +2263,43 @@ export const MODELS = [
         // publishes no cutoff.
         knowledgeCutoff: '2025-06-01',
     },
+    {
+        // The high-speed serving tier of the GLM-5.3-Flash series (~200 tokens/s
+        // per the GLM-5.3-Flash guide) — same series surface, higher rate card.
+        // Same 5.3 generation as glm-5.3-flash, so neither supersedes the other.
+        id: 'glm-5.3-flashx',
+        provider: 'zhipu',
+        label: 'GLM-5.3 FlashX',
+        description: 'High-speed native multimodal coding + agents — 1M context',
+        contextWindow: 1_048_576,
+        // "GLM-5.3-Flash series supports a maximum output length of 128K" —
+        // chat-completion API reference, checked 2026-09-23.
+        maxOutputTokens: 131_072,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        thinkingConfigurable: true,
+        // Same surface as glm-5.3-flash: low|high|max only, default max, thinking
+        // cannot be disabled. high is the balanced tier we default to.
+        supportedEffortLevels: ['low', 'high', 'max'],
+        defaultEffortLevel: 'high',
+        // Series input: text, images, video, files (glm-5.3-flashx appears in the
+        // chat-completion reference's image examples).
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        webSearchToolType: 'web_search',
+        // List card, verified 2026-09-23 on the Z.ai pricing page.
+        inputPricePerMTok: 0.37,
+        outputPricePerMTok: 1.25,
+        // GLM context cache: read ≈0.2× input, no write premium (storage free).
+        cacheReadPricePerMTok: 0.075,
+        cacheWritePricePerMTok: 0.37,
+        // Native host only: api.deepinfra.com/models/zai-org/GLM-5.3-FlashX is a
+        // 404 (checked 2026-09-23), so there is no US re-host to map.
+        regions: ['cn'],
+        // Not published by Z.ai — best-effort estimate for the 5.3 generation.
+        knowledgeCutoff: '2025-06-01',
+    },
     {
         id: 'glm-5.3-flash',
         provider: 'zhipu',

package/dist/types.d.ts

@@ -131,6 +131,31 @@ export interface ModelDefinition {
      * both together; until then, working-without-reasoning beats 400.
      */
     toolsRequireReasoningOff?: boolean;
+    /**
+     * The provider rejects a FORCED tool choice for this model — Anthropic
+     * `tool_choice` `any` / `tool` answer 400 on claude-fable-5-1 and
+     * claude-opus-5-5, and Moonshot rejects `required` while thinking is on,
+     * which it always is on kimi-k2.7-code (all probed live 2026-09-23). `auto`
+     * works and the model still calls the tool. Callers that would force a tool call (Synthase
+     * discovery, starting-point selection) must send `auto` for these models;
+     * molecule-dev resolves that in one place (`toolChoiceParam` in
+     * request-shape.ts) and its live dispatch check probes the forced shape for
+     * every model, so a new model with this restriction fails the check instead
+     * of every discovery turn.
+     */
+    rejectsForcedToolChoice?: boolean;
+    /**
+     * The provider rejects a caller-chosen `temperature` for this model —
+     * Anthropic answers 400 "`temperature` is deprecated for this model" on the
+     * Claude 5 family and Opus 4.7/4.8 (Opus 4.6, Sonnet 4.6 and Haiku 4.5 still
+     * accept it), and Moonshot's native host allows only 1 on kimi-k3 and
+     * kimi-k2.7-code (all probed live 2026-09-23). Omitting it is always safe. Callers that set a temperature (commit
+     * messages, starting-point selection) must omit it for these models;
+     * molecule-dev resolves that in one place (`temperatureParam` in
+     * request-shape.ts), and its live dispatch check sends the tool-less,
+     * temperature-0 shape to every model.
+     */
+    rejectsTemperature?: boolean;
     /**
      * Provider-specific server tool type for web search (e.g. `'web_search_20250305'`).
      * When set, the chat handler sends this as a ServerTool alongside custom tools.
@@ -234,6 +259,20 @@ export interface ModelDefinition {
      * over-bill this field exists to prevent. A wrapping window belongs to the
      * day it STARTS on, so its post-midnight tail is still matched against the
      * previous day.
+     *
+     * `excludedDatesUtc` lists `YYYY-MM-DD` dates on which no window applies —
+     * a provider's public holidays. DeepSeek's peak excludes Chinese public
+     * holidays (its pricing page, 2026-09-23); without the list those days bill
+     * at 2× while the provider charges off-peak. A date is matched against the
+     * day a window STARTS on (UTC), like `daysOfWeekUtc`. The list is DATA that
+     * runs out: check-model-freshness warns when the coming year has no dates.
+     *
+     * `rule` is the provider's own sentence defining these windows, verbatim,
+     * and the page that publishes it. check-model-freshness re-reads the page on
+     * every run and warns the moment the sentence changes — the windows are only
+     * as right as the last reading, and DeepSeek has amended its rule twice
+     * (a weekday qualifier by 2026-08-31, a holiday exclusion by 2026-09-23)
+     * without anything here noticing.
      */
     peakPricing?: {
         windows: {
@@ -242,6 +281,11 @@ export interface ModelDefinition {
             daysOfWeekUtc?: number[];
         }[];
         multiplier: number;
+        excludedDatesUtc?: string[];
+        rule?: {
+            url: string;
+            text: string;
+        };
     };
     /**
      * A price change the provider has ANNOUNCED with a dated effective instant,
@@ -294,6 +338,11 @@ export interface ModelDefinition {
                 daysOfWeekUtc?: number[];
             }[];
             multiplier: number;
+            excludedDatesUtc?: string[];
+            rule?: {
+                url: string;
+                text: string;
+            };
         };
         /** Where the change was announced, for the re-verify pass after it lands. */
         source?: string;

package/dist/types.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH;;;;;GAKG;AACH,MAAM,MAAM,YAAY,GACpB,WAAW,GACX,QAAQ,GACR,QAAQ,GACR,KAAK,GACL,UAAU,GACV,MAAM,GACN,UAAU,GACV,SAAS,GACT,SAAS,GACT,OAAO;AACT;;;;;GAKG;GACD,QAAQ,CAAA;AAEZ;;;;;;;;;;;;;GAaG;AACH,MAAM,MAAM,WAAW,GAAG,MAAM,CAAA;AAEhC;;;GAGG;AACH,MAAM,WAAW,eAAe;IAC9B,8DAA8D;IAC9D,EAAE,EAAE,MAAM,CAAA;IACV,2CAA2C;IAC3C,QAAQ,EAAE,YAAY,CAAA;IACtB,yDAAyD;IACzD,KAAK,EAAE,MAAM,CAAA;IACb,wCAAwC;IACxC,WAAW,EAAE,MAAM,CAAA;IACnB,8CAA8C;IAC9C,aAAa,EAAE,MAAM,CAAA;IACrB,0CAA0C;IAC1C,eAAe,EAAE,MAAM,CAAA;IACvB,uEAAuE;IACvE,gBAAgB,EAAE,OAAO,CAAA;IACzB,yFAAyF;IACzF,oBAAoB,EAAE,MAAM,CAAA;IAC5B;;;OAGG;IACH,oBAAoB,EAAE,OAAO,CAAA;IAC7B;;;;;;;;;;;;;;;;;;;OAmBG;IACH,qBAAqB,CAAC,EAAE,WAAW,EAAE,CAAA;IACrC;;;;OAIG;IACH,kBAAkB,CAAC,EAAE,WAAW,CAAA;IAChC;;;;;;;;;;;OAWG;IACH,kBAAkB,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAC3C,mEAAmE;IACnE,cAAc,EAAE,OAAO,CAAA;IACvB,iDAAiD;IACjD,qBAAqB,EAAE,OAAO,CAAA;IAC9B,8DAA8D;IAC9D,aAAa,EAAE,OAAO,CAAA;IACtB;;;;;;;;;;;;;;;;;;;OAmBG;IACH,wBAAwB,CAAC,EAAE,OAAO,CAAA;IAClC;;;;OAIG;IACH,iBAAiB,CAAC,EAAE,MAAM,CAAA;IAC1B;;;OAGG;IACH,qBAAqB,CAAC,EAAE,MAAM,CAAA;IAC9B;;;OAGG;IACH,gBAAgB,CAAC,EAAE,MAAM,CAAA;IACzB,wFAAwF;IACxF,QAAQ,CAAC,EAAE,OAAO,CAAA;IAClB;;;;;;OAMG;IACH,eAAe,CAAC,EAAE,MAAM,EAAE,CAAA;IAC1B;;;;;;;OAOG;IACH,OAAO,CAAC,EAAE,MAAM,EAAE,CAAA;IAClB;;;;;;;;OAQG;IACH,aAAa,CAAC,EAAE,MAAM,CACpB,MAAM,EACN;QACE,6DAA6D;QAC7D,iBAAiB,EAAE,MAAM,CAAA;QACzB,qDAAqD;QACrD,kBAAkB,EAAE,MAAM,CAAA;QAC1B,gEAAgE;QAChE,qBAAqB,CAAC,EAAE,MAAM,CAAA;QAC9B,iEAAiE;QACjE,sBAAsB,CAAC,EAAE,MAAM,CAAA;KAChC,CACF,CAAA;IACD,sEAAsE;IACtE,iBAAiB,EAAE,MAAM,CAAA;IACzB,8CAA8C;IAC9C,kBAAkB,EAAE,MAAM,CAAA;IAC1B;;;;;;;;;;;OAWG;IACH,qBAAqB,EAAE,MAAM,CAAA;IAC7B;;;;;;;;;;OAUG;IACH,sBAAsB,EAAE,MAAM,CAAA;IAC9B;;;;;;;;;;;;;;;;;;;;OAoBG;IACH,WAAW,CAAC,EAAE;QACZ,OAAO,EAAE;YAAE,cAAc,EAAE,MAAM,CAAC;YAAC,YAAY,EAAE,MAAM,CAAC;YAAC,aAAa,CAAC,EAAE,MAAM,EAAE,CAAA;SAAE,EAAE,CAAA;QACrF,UAAU,EAAE,MAAM,CAAA;KACnB,CAAA;IACD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;OA+BG;IACH,gBAAgB,CAAC,EAAE;QACjB,sDAAsD;QACtD,aAAa,EAAE,MAAM,CAAA;QACrB,8EAA8E;QAC9E,iBAAiB,EAAE,MAAM,CAAA;QACzB,oEAAoE;QACpE,kBAAkB,EAAE,MAAM,CAAA;QAC1B,iFAAiF;QACjF,qBAAqB,EAAE,MAAM,CAAA;QAC7B,kFAAkF;QAClF,sBAAsB,EAAE,MAAM,CAAA;QAC9B,yFAAyF;QACzF,WAAW,CAAC,EAAE;YACZ,OAAO,EAAE;gBAAE,cAAc,EAAE,MAAM,CAAC;gBAAC,YAAY,EAAE,MAAM,CAAC;gBAAC,aAAa,CAAC,EAAE,MAAM,EAAE,CAAA;aAAE,EAAE,CAAA;YACrF,UAAU,EAAE,MAAM,CAAA;SACnB,CAAA;QACD,6EAA6E;QAC7E,MAAM,CAAC,EAAE,MAAM,CAAA;KAChB,CAAA;IACD;;;;;;;;;;OAUG;IACH,WAAW,CAAC,EAAE;QACZ,gEAAgE;QAChE,iBAAiB,EAAE,MAAM,CAAA;QACzB,wDAAwD;QACxD,kBAAkB,EAAE,MAAM,CAAA;QAC1B,mEAAmE;QACnE,qBAAqB,EAAE,MAAM,CAAA;QAC7B,oEAAoE;QACpE,sBAAsB,EAAE,MAAM,CAAA;KAC/B,CAAA;IACD,mDAAmD;IACnD,eAAe,EAAE,MAAM,CAAA;IACvB;;;;;;;;;OASG;IACH,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB;;;;;;;;;;;;;;OAcG;IACH,QAAQ,CAAC,EAAE,OAAO,CAAA;IAClB;;;;;;;;;;;;;;;;;;;;;;;;;;;OA2BG;IACH,YAAY,CAAC,EAAE,MAAM,CAAA;CACtB;AAED;;;;;;GAMG;AACH,MAAM,WAAW,iBAAiB;IAChC,6DAA6D;IAC7D,IAAI,EAAE,MAAM,CAAA;IACZ,gEAAgE;IAChE,OAAO,EAAE,MAAM,CAAA;IACf,8EAA8E;IAC9E,MAAM,EAAE,MAAM,CAAA;IACd,4EAA4E;IAC5E,OAAO,EAAE,MAAM,CAAA;CAChB;AAED;;GAEG;AACH,MAAM,WAAW,kBAAkB;IACjC,MAAM,EAAE,eAAe,EAAE,CAAA;IACzB;;;;OAIG;IACH,QAAQ,CAAC,EAAE,iBAAiB,CAAA;CAC7B"}
+{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH;;;;;GAKG;AACH,MAAM,MAAM,YAAY,GACpB,WAAW,GACX,QAAQ,GACR,QAAQ,GACR,KAAK,GACL,UAAU,GACV,MAAM,GACN,UAAU,GACV,SAAS,GACT,SAAS,GACT,OAAO;AACT;;;;;GAKG;GACD,QAAQ,CAAA;AAEZ;;;;;;;;;;;;;GAaG;AACH,MAAM,MAAM,WAAW,GAAG,MAAM,CAAA;AAEhC;;;GAGG;AACH,MAAM,WAAW,eAAe;IAC9B,8DAA8D;IAC9D,EAAE,EAAE,MAAM,CAAA;IACV,2CAA2C;IAC3C,QAAQ,EAAE,YAAY,CAAA;IACtB,yDAAyD;IACzD,KAAK,EAAE,MAAM,CAAA;IACb,wCAAwC;IACxC,WAAW,EAAE,MAAM,CAAA;IACnB,8CAA8C;IAC9C,aAAa,EAAE,MAAM,CAAA;IACrB,0CAA0C;IAC1C,eAAe,EAAE,MAAM,CAAA;IACvB,uEAAuE;IACvE,gBAAgB,EAAE,OAAO,CAAA;IACzB,yFAAyF;IACzF,oBAAoB,EAAE,MAAM,CAAA;IAC5B;;;OAGG;IACH,oBAAoB,EAAE,OAAO,CAAA;IAC7B;;;;;;;;;;;;;;;;;;;OAmBG;IACH,qBAAqB,CAAC,EAAE,WAAW,EAAE,CAAA;IACrC;;;;OAIG;IACH,kBAAkB,CAAC,EAAE,WAAW,CAAA;IAChC;;;;;;;;;;;OAWG;IACH,kBAAkB,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAC3C,mEAAmE;IACnE,cAAc,EAAE,OAAO,CAAA;IACvB,iDAAiD;IACjD,qBAAqB,EAAE,OAAO,CAAA;IAC9B,8DAA8D;IAC9D,aAAa,EAAE,OAAO,CAAA;IACtB;;;;;;;;;;;;;;;;;;;OAmBG;IACH,wBAAwB,CAAC,EAAE,OAAO,CAAA;IAClC;;;;;;;;;;;OAWG;IACH,uBAAuB,CAAC,EAAE,OAAO,CAAA;IACjC;;;;;;;;;;OAUG;IACH,kBAAkB,CAAC,EAAE,OAAO,CAAA;IAC5B;;;;OAIG;IACH,iBAAiB,CAAC,EAAE,MAAM,CAAA;IAC1B;;;OAGG;IACH,qBAAqB,CAAC,EAAE,MAAM,CAAA;IAC9B;;;OAGG;IACH,gBAAgB,CAAC,EAAE,MAAM,CAAA;IACzB,wFAAwF;IACxF,QAAQ,CAAC,EAAE,OAAO,CAAA;IAClB;;;;;;OAMG;IACH,eAAe,CAAC,EAAE,MAAM,EAAE,CAAA;IAC1B;;;;;;;OAOG;IACH,OAAO,CAAC,EAAE,MAAM,EAAE,CAAA;IAClB;;;;;;;;OAQG;IACH,aAAa,CAAC,EAAE,MAAM,CACpB,MAAM,EACN;QACE,6DAA6D;QAC7D,iBAAiB,EAAE,MAAM,CAAA;QACzB,qDAAqD;QACrD,kBAAkB,EAAE,MAAM,CAAA;QAC1B,gEAAgE;QAChE,qBAAqB,CAAC,EAAE,MAAM,CAAA;QAC9B,iEAAiE;QACjE,sBAAsB,CAAC,EAAE,MAAM,CAAA;KAChC,CACF,CAAA;IACD,sEAAsE;IACtE,iBAAiB,EAAE,MAAM,CAAA;IACzB,8CAA8C;IAC9C,kBAAkB,EAAE,MAAM,CAAA;IAC1B;;;;;;;;;;;OAWG;IACH,qBAAqB,EAAE,MAAM,CAAA;IAC7B;;;;;;;;;;OAUG;IACH,sBAAsB,EAAE,MAAM,CAAA;IAC9B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;OAkCG;IACH,WAAW,CAAC,EAAE;QACZ,OAAO,EAAE;YAAE,cAAc,EAAE,MAAM,CAAC;YAAC,YAAY,EAAE,MAAM,CAAC;YAAC,aAAa,CAAC,EAAE,MAAM,EAAE,CAAA;SAAE,EAAE,CAAA;QACrF,UAAU,EAAE,MAAM,CAAA;QAClB,gBAAgB,CAAC,EAAE,MAAM,EAAE,CAAA;QAC3B,IAAI,CAAC,EAAE;YAAE,GAAG,EAAE,MAAM,CAAC;YAAC,IAAI,EAAE,MAAM,CAAA;SAAE,CAAA;KACrC,CAAA;IACD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;OA+BG;IACH,gBAAgB,CAAC,EAAE;QACjB,sDAAsD;QACtD,aAAa,EAAE,MAAM,CAAA;QACrB,8EAA8E;QAC9E,iBAAiB,EAAE,MAAM,CAAA;QACzB,oEAAoE;QACpE,kBAAkB,EAAE,MAAM,CAAA;QAC1B,iFAAiF;QACjF,qBAAqB,EAAE,MAAM,CAAA;QAC7B,kFAAkF;QAClF,sBAAsB,EAAE,MAAM,CAAA;QAC9B,yFAAyF;QACzF,WAAW,CAAC,EAAE;YACZ,OAAO,EAAE;gBAAE,cAAc,EAAE,MAAM,CAAC;gBAAC,YAAY,EAAE,MAAM,CAAC;gBAAC,aAAa,CAAC,EAAE,MAAM,EAAE,CAAA;aAAE,EAAE,CAAA;YACrF,UAAU,EAAE,MAAM,CAAA;YAClB,gBAAgB,CAAC,EAAE,MAAM,EAAE,CAAA;YAC3B,IAAI,CAAC,EAAE;gBAAE,GAAG,EAAE,MAAM,CAAC;gBAAC,IAAI,EAAE,MAAM,CAAA;aAAE,CAAA;SACrC,CAAA;QACD,6EAA6E;QAC7E,MAAM,CAAC,EAAE,MAAM,CAAA;KAChB,CAAA;IACD;;;;;;;;;;OAUG;IACH,WAAW,CAAC,EAAE;QACZ,gEAAgE;QAChE,iBAAiB,EAAE,MAAM,CAAA;QACzB,wDAAwD;QACxD,kBAAkB,EAAE,MAAM,CAAA;QAC1B,mEAAmE;QACnE,qBAAqB,EAAE,MAAM,CAAA;QAC7B,oEAAoE;QACpE,sBAAsB,EAAE,MAAM,CAAA;KAC/B,CAAA;IACD,mDAAmD;IACnD,eAAe,EAAE,MAAM,CAAA;IACvB;;;;;;;;;OASG;IACH,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB;;;;;;;;;;;;;;OAcG;IACH,QAAQ,CAAC,EAAE,OAAO,CAAA;IAClB;;;;;;;;;;;;;;;;;;;;;;;;;;;OA2BG;IACH,YAAY,CAAC,EAAE,MAAM,CAAA;CACtB;AAED;;;;;;GAMG;AACH,MAAM,WAAW,iBAAiB;IAChC,6DAA6D;IAC7D,IAAI,EAAE,MAAM,CAAA;IACZ,gEAAgE;IAChE,OAAO,EAAE,MAAM,CAAA;IACf,8EAA8E;IAC9E,MAAM,EAAE,MAAM,CAAA;IACd,4EAA4E;IAC5E,OAAO,EAAE,MAAM,CAAA;CAChB;AAED;;GAEG;AACH,MAAM,WAAW,kBAAkB;IACjC,MAAM,EAAE,eAAe,EAAE,CAAA;IACzB;;;;OAIG;IACH,QAAQ,CAAC,EAAE,iBAAiB,CAAA;CAC7B"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@molecule/api-resource-ai-models",
-  "version": "1.6.3",
+  "version": "1.8.0",
   "description": "AI model catalog — server-side source of truth plus an authentication-gated discovery endpoint",
   "type": "module",
   "main": "dist/index.js",