@molecule/api-resource-ai-models 1.6.3 → 1.7.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +84 -3
- package/dist/lookup.d.ts.map +1 -1
- package/dist/lookup.js +10 -4
- package/dist/models.d.ts +38 -2
- package/dist/models.d.ts.map +1 -1
- package/dist/models.js +325 -40
- package/dist/types.d.ts +49 -0
- package/dist/types.d.ts.map +1 -1
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -3,7 +3,7 @@ AUTO-GENERATED — DO NOT EDIT THIS FILE.
|
|
|
3
3
|
Generated by `mlcl sync-docs` from the package's src/index.ts JSDoc + mlcl/registry.json.
|
|
4
4
|
Edits here are overwritten on the next commit (molecule's pre-commit hook regenerates).
|
|
5
5
|
To change this document, edit the module-level JSDoc in src/index.ts.
|
|
6
|
-
Generated: 2026-09-
|
|
6
|
+
Generated: 2026-09-23T14:26:31.162Z
|
|
7
7
|
-->
|
|
8
8
|
|
|
9
9
|
# @molecule/api-resource-ai-models
|
|
@@ -159,6 +159,31 @@ interface ModelDefinition {
|
|
|
159
159
|
* both together; until then, working-without-reasoning beats 400.
|
|
160
160
|
*/
|
|
161
161
|
toolsRequireReasoningOff?: boolean
|
|
162
|
+
/**
|
|
163
|
+
* The provider rejects a FORCED tool choice for this model — Anthropic
|
|
164
|
+
* `tool_choice` `any` / `tool` answer 400 on claude-fable-5-1 and
|
|
165
|
+
* claude-opus-5-5, and Moonshot rejects `required` while thinking is on,
|
|
166
|
+
* which it always is on kimi-k2.7-code (all probed live 2026-09-23). `auto`
|
|
167
|
+
* works and the model still calls the tool. Callers that would force a tool call (Synthase
|
|
168
|
+
* discovery, starting-point selection) must send `auto` for these models;
|
|
169
|
+
* molecule-dev resolves that in one place (`toolChoiceParam` in
|
|
170
|
+
* request-shape.ts) and its live dispatch check probes the forced shape for
|
|
171
|
+
* every model, so a new model with this restriction fails the check instead
|
|
172
|
+
* of every discovery turn.
|
|
173
|
+
*/
|
|
174
|
+
rejectsForcedToolChoice?: boolean
|
|
175
|
+
/**
|
|
176
|
+
* The provider rejects a caller-chosen `temperature` for this model —
|
|
177
|
+
* Anthropic answers 400 "`temperature` is deprecated for this model" on the
|
|
178
|
+
* Claude 5 family and Opus 4.7/4.8 (Opus 4.6, Sonnet 4.6 and Haiku 4.5 still
|
|
179
|
+
* accept it), and Moonshot's native host allows only 1 on kimi-k3 and
|
|
180
|
+
* kimi-k2.7-code (all probed live 2026-09-23). Omitting it is always safe. Callers that set a temperature (commit
|
|
181
|
+
* messages, starting-point selection) must omit it for these models;
|
|
182
|
+
* molecule-dev resolves that in one place (`temperatureParam` in
|
|
183
|
+
* request-shape.ts), and its live dispatch check sends the tool-less,
|
|
184
|
+
* temperature-0 shape to every model.
|
|
185
|
+
*/
|
|
186
|
+
rejectsTemperature?: boolean
|
|
162
187
|
/**
|
|
163
188
|
* Provider-specific server tool type for web search (e.g. `'web_search_20250305'`).
|
|
164
189
|
* When set, the chat handler sends this as a ServerTool alongside custom tools.
|
|
@@ -265,10 +290,26 @@ interface ModelDefinition {
|
|
|
265
290
|
* over-bill this field exists to prevent. A wrapping window belongs to the
|
|
266
291
|
* day it STARTS on, so its post-midnight tail is still matched against the
|
|
267
292
|
* previous day.
|
|
293
|
+
*
|
|
294
|
+
* `excludedDatesUtc` lists `YYYY-MM-DD` dates on which no window applies —
|
|
295
|
+
* a provider's public holidays. DeepSeek's peak excludes Chinese public
|
|
296
|
+
* holidays (its pricing page, 2026-09-23); without the list those days bill
|
|
297
|
+
* at 2× while the provider charges off-peak. A date is matched against the
|
|
298
|
+
* day a window STARTS on (UTC), like `daysOfWeekUtc`. The list is DATA that
|
|
299
|
+
* runs out: check-model-freshness warns when the coming year has no dates.
|
|
300
|
+
*
|
|
301
|
+
* `rule` is the provider's own sentence defining these windows, verbatim,
|
|
302
|
+
* and the page that publishes it. check-model-freshness re-reads the page on
|
|
303
|
+
* every run and warns the moment the sentence changes — the windows are only
|
|
304
|
+
* as right as the last reading, and DeepSeek has amended its rule twice
|
|
305
|
+
* (a weekday qualifier by 2026-08-31, a holiday exclusion by 2026-09-23)
|
|
306
|
+
* without anything here noticing.
|
|
268
307
|
*/
|
|
269
308
|
peakPricing?: {
|
|
270
309
|
windows: { startMinuteUtc: number; endMinuteUtc: number; daysOfWeekUtc?: number[] }[]
|
|
271
310
|
multiplier: number
|
|
311
|
+
excludedDatesUtc?: string[]
|
|
312
|
+
rule?: { url: string; text: string }
|
|
272
313
|
}
|
|
273
314
|
/**
|
|
274
315
|
* A price change the provider has ANNOUNCED with a dated effective instant,
|
|
@@ -317,6 +358,8 @@ interface ModelDefinition {
|
|
|
317
358
|
peakPricing?: {
|
|
318
359
|
windows: { startMinuteUtc: number; endMinuteUtc: number; daysOfWeekUtc?: number[] }[]
|
|
319
360
|
multiplier: number
|
|
361
|
+
excludedDatesUtc?: string[]
|
|
362
|
+
rule?: { url: string; text: string }
|
|
320
363
|
}
|
|
321
364
|
/** Where the change was announced, for the re-verify pass after it lands. */
|
|
322
365
|
source?: string
|
|
@@ -543,6 +586,8 @@ function effectivePeakPricing(
|
|
|
543
586
|
| {
|
|
544
587
|
windows: { startMinuteUtc: number; endMinuteUtc: number; daysOfWeekUtc?: number[] }[]
|
|
545
588
|
multiplier: number
|
|
589
|
+
excludedDatesUtc?: string[]
|
|
590
|
+
rule?: { url: string; text: string }
|
|
546
591
|
}
|
|
547
592
|
| undefined
|
|
548
593
|
```
|
|
@@ -786,7 +831,8 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
|
|
|
786
831
|
`npm run check:model-freshness` from the workspace root):
|
|
787
832
|
|
|
788
833
|
- Anthropic: https://platform.claude.com/docs/en/about-claude/models/overview
|
|
789
|
-
- /docs/en/build-with-claude/effort (fable-5 / opus-5 / sonnet-5
|
|
834
|
+
- /docs/en/build-with-claude/effort (fable-5-1 / opus-5-5 / sonnet-5
|
|
835
|
+
current as of 2026-09-23 — see the dated notes below; historically fable-5 / opus-5 / sonnet-5 current;
|
|
790
836
|
opus-4-8 superseded by opus-5 at identical pricing but still served — it is
|
|
791
837
|
the recommended refusal-fallback model; effort ladder on all three current
|
|
792
838
|
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
|
|
|
816
862
|
any/tool now 400s, thinking blocks are model-bound, editing earlier turns
|
|
817
863
|
invalidates them) — Synthase sends tool_choice auto and is append-only,
|
|
818
864
|
and the pre-commit dispatch probe covers the entry as sent.)
|
|
865
|
+
(verified 2026-09-23 — ADDED claude-opus-5-5, released 2026-09-22 and
|
|
866
|
+
listed as "Active (latest)" on /docs/en/models/opus-5-5/overview; the
|
|
867
|
+
overview page now lists claude-opus-5 under "Legacy models (still
|
|
868
|
+
available)" → superseded, with opus-4-8/4-7/4-6 repointed one hop.
|
|
869
|
+
Pricing page: $4/$20, 5m cache write $5, cache hits $0.20 (footnote 2:
|
|
870
|
+
0.05× input on Opus 5.5 only), fast mode $8/$40. 1M ctx / 128K out, text +
|
|
871
|
+
image input, reliable knowledge cutoff Jun 2026. Effort page: all five
|
|
872
|
+
levels, default MEDIUM. What's-new page: thinking always on
|
|
873
|
+
(disabled/budget_tokens 400), tool_choice any/tool 400, computer_20251124
|
|
874
|
+
400 — Synthase sends none of those. Every other Anthropic price on the
|
|
875
|
+
pricing page is unchanged.)
|
|
819
876
|
- OpenAI: https://developers.openai.com/api/docs/pricing (GPT-5.6 family GA
|
|
820
877
|
2026-07-09; REPRICED 2026-07-30: -luna cut 80% to $0.20/$1.20, -terra cut
|
|
821
878
|
20% to $2/$12, -sol unchanged $5/$30; cache read 0.1× input; gpt-5.5/
|
|
@@ -840,6 +897,14 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
|
|
|
840
897
|
> in the catalog because it cannot serve a tool-carrying request on the
|
|
841
898
|
> bond's /v1/chat/completions endpoint; see the DO NOT ADD block above the
|
|
842
899
|
> OpenAI entries for the live 400s and the condition that lifts it.)
|
|
900
|
+
> (re-verified 2026-09-23: ADDED gpt-6-sol ($2/$10, cached $0.20, cache
|
|
901
|
+
writes $2.50) and gpt-6-luna ($0.10/$0.50, cached $0.01, cache writes
|
|
902
|
+
$0.125), released 2026-09-22 — both with the same unmodeled >272K band,
|
|
903
|
+
> 1.05M ctx / 128K out, image input, knowledge cutoff Apr 20 / May 18 2026
|
|
904
|
+
> per their model pages. Unlike astra they serve tool-carrying requests on
|
|
905
|
+
> /v1/chat/completions with reasoning_effort 'none'. The 5.6 family is now
|
|
906
|
+
> superseded; every 5.6 price on the page is unchanged and the -sol promo
|
|
907
|
+
> footnote still reads "at least through November 21, 2026".)
|
|
843
908
|
- Google: https://ai.google.dev/gemini-api/docs/pricing (gemini-3.6-flash GA
|
|
844
909
|
2026-07-21 $1.50/$7.50 supersedes 3.5-flash as the agentic flagship;
|
|
845
910
|
gemini-3.1-pro-preview still the pro tier — "3.5 Pro" has NOT shipped as
|
|
@@ -872,8 +937,16 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
|
|
|
872
937
|
not modeled; reasoning_effort low|medium|high default high, image input;
|
|
873
938
|
grok-4.3 still served at $1.25/$2.50 with the bigger 1M window;
|
|
874
939
|
grok-code-fast-1 no longer listed — retires 2026-08-15)
|
|
940
|
+
- Chinese public holidays (DeepSeek's peak excludes them):
|
|
941
|
+
https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html
|
|
942
|
+
(the State Council's 2026 notice; see CHINA_PUBLIC_HOLIDAYS)
|
|
875
943
|
- DeepSeek: https://api-docs.deepseek.com/quick_start/pricing +
|
|
876
|
-
/updates/ (verified 2026-09-
|
|
944
|
+
/updates/ (verified 2026-09-23 — the announced 2026-09-14 routing of
|
|
945
|
+
`deepseek-v4-pro` to V4.1 Flash was WITHDRAWN: V4 Pro keeps its card
|
|
946
|
+
0.66/1.98/0.022 off-peak, so its staged `scheduledPricing` was deleted.
|
|
947
|
+
The card now also says peak hours exclude Chinese public holidays, which
|
|
948
|
+
`peakPricing` cannot express — peak is billed on those weekdays too.
|
|
949
|
+
Earlier, 2026-09-10: legacy deepseek-chat/-reasoner ids fully
|
|
877
950
|
retired 2026-07-24 — never in this catalog. The 2026-09-10 re-read caught
|
|
878
951
|
the V4.1-Flash release DAY-OF: new evergreen id `deepseek-flash` at
|
|
879
952
|
off-peak miss $0.15 / hit $0.003 / out $0.6 (peak ×2, same Mon-Fri UTC
|
|
@@ -958,6 +1031,14 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
|
|
|
958
1031
|
card again — $0.15/$0.50, cached input $0.03, no promo footnote. models.dev
|
|
959
1032
|
still carries the expired promo rate, so the KNOWN_DIVERGENCES entry now
|
|
960
1033
|
runs to a fixed 2026-12-09 re-verify date.)
|
|
1034
|
+
(verified 2026-09-23: glm-5.3-flashx ADDED — the ~200 tokens/s tier of the
|
|
1035
|
+
GLM-5.3-Flash series. Pricing page: $0.37/$1.25, cached input $0.075,
|
|
1036
|
+
cache storage free. Chat-completion reference: id `glm-5.3-flashx` in the
|
|
1037
|
+
vision request enum, 128K max output, reasoning_effort low|high|max
|
|
1038
|
+
(default max), function tools supported. GLM-5.3-Flash guide: 1M ctx,
|
|
1039
|
+
video/image/text/file input. No published knowledge cutoff. Not on
|
|
1040
|
+
DeepInfra (zai-org/GLM-5.3-FlashX 404s), so cn-region-only. glm-5.3 and
|
|
1041
|
+
glm-5.3-flash rates unchanged on the same page.)
|
|
961
1042
|
|
|
962
1043
|
Knowledge-cutoff dates on non-Anthropic entries are best-effort estimates
|
|
963
1044
|
where the provider doesn't publish one; the provider sources above verify
|
package/dist/lookup.d.ts.map
CHANGED
|
@@ -1 +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,
|
|
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,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
CHANGED
|
@@ -217,14 +217,20 @@ export function priceMultiplierAt(modelDef, at, region) {
|
|
|
217
217
|
: minute >= w.startMinuteUtc && minute < w.endMinuteUtc;
|
|
218
218
|
if (!inWindow)
|
|
219
219
|
continue;
|
|
220
|
+
// A wrapping window belongs to the day it STARTED on, so its
|
|
221
|
+
// post-midnight tail is matched against the previous UTC day — a
|
|
222
|
+
// Friday-only 23:00-02:00 window must still be peak at Saturday 00:30.
|
|
223
|
+
const startedYesterday = wraps && minute < w.endMinuteUtc;
|
|
220
224
|
if (w.daysOfWeekUtc && w.daysOfWeekUtc.length > 0) {
|
|
221
|
-
|
|
222
|
-
// post-midnight tail is matched against the previous UTC day — a
|
|
223
|
-
// Friday-only 23:00-02:00 window must still be peak at Saturday 00:30.
|
|
224
|
-
const day = wraps && minute < w.endMinuteUtc ? (at.getUTCDay() + 6) % 7 : at.getUTCDay();
|
|
225
|
+
const day = startedYesterday ? (at.getUTCDay() + 6) % 7 : at.getUTCDay();
|
|
225
226
|
if (!w.daysOfWeekUtc.includes(day))
|
|
226
227
|
continue;
|
|
227
228
|
}
|
|
229
|
+
if (peak.excludedDatesUtc && peak.excludedDatesUtc.length > 0) {
|
|
230
|
+
const started = new Date(at.getTime() - (startedYesterday ? 86_400_000 : 0));
|
|
231
|
+
if (peak.excludedDatesUtc.includes(started.toISOString().slice(0, 10)))
|
|
232
|
+
continue;
|
|
233
|
+
}
|
|
228
234
|
return peak.multiplier;
|
|
229
235
|
}
|
|
230
236
|
return 1;
|
package/dist/models.d.ts
CHANGED
|
@@ -54,7 +54,8 @@ import type { ModelDefinition } from './types.js';
|
|
|
54
54
|
* 2026-07-30 GPT-5.6 repricing — cross-check prices against models.dev with
|
|
55
55
|
* `npm run check:model-freshness` from the workspace root):
|
|
56
56
|
* - Anthropic: https://platform.claude.com/docs/en/about-claude/models/overview
|
|
57
|
-
* + /docs/en/build-with-claude/effort (fable-5 / opus-5 / sonnet-5
|
|
57
|
+
* + /docs/en/build-with-claude/effort (fable-5-1 / opus-5-5 / sonnet-5
|
|
58
|
+
* current as of 2026-09-23 — see the dated notes below; historically fable-5 / opus-5 / sonnet-5 current;
|
|
58
59
|
* opus-4-8 superseded by opus-5 at identical pricing but still served — it is
|
|
59
60
|
* the recommended refusal-fallback model; effort ladder on all three current
|
|
60
61
|
* models is low|medium|high|xhigh|max; budget_tokens 400s on 4.7+)
|
|
@@ -84,6 +85,17 @@ import type { ModelDefinition } from './types.js';
|
|
|
84
85
|
* any/tool now 400s, thinking blocks are model-bound, editing earlier turns
|
|
85
86
|
* invalidates them) — Synthase sends tool_choice auto and is append-only,
|
|
86
87
|
* and the pre-commit dispatch probe covers the entry as sent.)
|
|
88
|
+
* (verified 2026-09-23 — ADDED claude-opus-5-5, released 2026-09-22 and
|
|
89
|
+
* listed as "Active (latest)" on /docs/en/models/opus-5-5/overview; the
|
|
90
|
+
* overview page now lists claude-opus-5 under "Legacy models (still
|
|
91
|
+
* available)" → superseded, with opus-4-8/4-7/4-6 repointed one hop.
|
|
92
|
+
* Pricing page: $4/$20, 5m cache write $5, cache hits $0.20 (footnote 2:
|
|
93
|
+
* 0.05× input on Opus 5.5 only), fast mode $8/$40. 1M ctx / 128K out, text +
|
|
94
|
+
* image input, reliable knowledge cutoff Jun 2026. Effort page: all five
|
|
95
|
+
* levels, default MEDIUM. What's-new page: thinking always on
|
|
96
|
+
* (disabled/budget_tokens 400), tool_choice any/tool 400, computer_20251124
|
|
97
|
+
* 400 — Synthase sends none of those. Every other Anthropic price on the
|
|
98
|
+
* pricing page is unchanged.)
|
|
87
99
|
* - OpenAI: https://developers.openai.com/api/docs/pricing (GPT-5.6 family GA
|
|
88
100
|
* 2026-07-09; REPRICED 2026-07-30: -luna cut 80% to $0.20/$1.20, -terra cut
|
|
89
101
|
* 20% to $2/$12, -sol unchanged $5/$30; cache read 0.1× input; gpt-5.5/
|
|
@@ -108,6 +120,14 @@ import type { ModelDefinition } from './types.js';
|
|
|
108
120
|
* in the catalog because it cannot serve a tool-carrying request on the
|
|
109
121
|
* bond's /v1/chat/completions endpoint; see the DO NOT ADD block above the
|
|
110
122
|
* OpenAI entries for the live 400s and the condition that lifts it.)
|
|
123
|
+
* (re-verified 2026-09-23: ADDED gpt-6-sol ($2/$10, cached $0.20, cache
|
|
124
|
+
* writes $2.50) and gpt-6-luna ($0.10/$0.50, cached $0.01, cache writes
|
|
125
|
+
* $0.125), released 2026-09-22 — both with the same unmodeled >272K band,
|
|
126
|
+
* 1.05M ctx / 128K out, image input, knowledge cutoff Apr 20 / May 18 2026
|
|
127
|
+
* per their model pages. Unlike astra they serve tool-carrying requests on
|
|
128
|
+
* /v1/chat/completions with reasoning_effort 'none'. The 5.6 family is now
|
|
129
|
+
* superseded; every 5.6 price on the page is unchanged and the -sol promo
|
|
130
|
+
* footnote still reads "at least through November 21, 2026".)
|
|
111
131
|
* - Google: https://ai.google.dev/gemini-api/docs/pricing (gemini-3.6-flash GA
|
|
112
132
|
* 2026-07-21 $1.50/$7.50 supersedes 3.5-flash as the agentic flagship;
|
|
113
133
|
* gemini-3.1-pro-preview still the pro tier — "3.5 Pro" has NOT shipped as
|
|
@@ -140,8 +160,16 @@ import type { ModelDefinition } from './types.js';
|
|
|
140
160
|
* not modeled; reasoning_effort low|medium|high default high, image input;
|
|
141
161
|
* grok-4.3 still served at $1.25/$2.50 with the bigger 1M window;
|
|
142
162
|
* grok-code-fast-1 no longer listed — retires 2026-08-15)
|
|
163
|
+
* - Chinese public holidays (DeepSeek's peak excludes them):
|
|
164
|
+
* https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html
|
|
165
|
+
* (the State Council's 2026 notice; see CHINA_PUBLIC_HOLIDAYS)
|
|
143
166
|
* - DeepSeek: https://api-docs.deepseek.com/quick_start/pricing +
|
|
144
|
-
* /updates/ (verified 2026-09-
|
|
167
|
+
* /updates/ (verified 2026-09-23 — the announced 2026-09-14 routing of
|
|
168
|
+
* `deepseek-v4-pro` to V4.1 Flash was WITHDRAWN: V4 Pro keeps its card
|
|
169
|
+
* 0.66/1.98/0.022 off-peak, so its staged `scheduledPricing` was deleted.
|
|
170
|
+
* The card now also says peak hours exclude Chinese public holidays, which
|
|
171
|
+
* `peakPricing` cannot express — peak is billed on those weekdays too.
|
|
172
|
+
* Earlier, 2026-09-10: legacy deepseek-chat/-reasoner ids fully
|
|
145
173
|
* retired 2026-07-24 — never in this catalog. The 2026-09-10 re-read caught
|
|
146
174
|
* the V4.1-Flash release DAY-OF: new evergreen id `deepseek-flash` at
|
|
147
175
|
* off-peak miss $0.15 / hit $0.003 / out $0.6 (peak ×2, same Mon-Fri UTC
|
|
@@ -226,6 +254,14 @@ import type { ModelDefinition } from './types.js';
|
|
|
226
254
|
* card again — $0.15/$0.50, cached input $0.03, no promo footnote. models.dev
|
|
227
255
|
* still carries the expired promo rate, so the KNOWN_DIVERGENCES entry now
|
|
228
256
|
* runs to a fixed 2026-12-09 re-verify date.)
|
|
257
|
+
* (verified 2026-09-23: glm-5.3-flashx ADDED — the ~200 tokens/s tier of the
|
|
258
|
+
* GLM-5.3-Flash series. Pricing page: $0.37/$1.25, cached input $0.075,
|
|
259
|
+
* cache storage free. Chat-completion reference: id `glm-5.3-flashx` in the
|
|
260
|
+
* vision request enum, 128K max output, reasoning_effort low|high|max
|
|
261
|
+
* (default max), function tools supported. GLM-5.3-Flash guide: 1M ctx,
|
|
262
|
+
* video/image/text/file input. No published knowledge cutoff. Not on
|
|
263
|
+
* DeepInfra (zai-org/GLM-5.3-FlashX 404s), so cn-region-only. glm-5.3 and
|
|
264
|
+
* glm-5.3-flash rates unchanged on the same page.)
|
|
229
265
|
*
|
|
230
266
|
* Knowledge-cutoff dates on non-Anthropic entries are best-effort estimates
|
|
231
267
|
* where the provider doesn't publish one; the provider sources above verify
|
package/dist/models.d.ts.map
CHANGED
|
@@ -1 +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;
|
|
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;AAwDjD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkQG;AACH,eAAO,MAAM,MAAM,EAAE,SAAS,eAAe,EA2jEnC,CAAA"}
|
package/dist/models.js
CHANGED
|
@@ -12,6 +12,51 @@
|
|
|
12
12
|
* prices by business day rather than by hour alone.
|
|
13
13
|
*/
|
|
14
14
|
const WEEKDAYS_UTC = [1, 2, 3, 4, 5];
|
|
15
|
+
/**
|
|
16
|
+
* DeepSeek's own sentence defining its peak windows, verbatim (pricing page,
|
|
17
|
+
* re-read 2026-09-23). check-model-freshness warns when the page stops saying
|
|
18
|
+
* exactly this — then the windows and CHINA_PUBLIC_HOLIDAYS need re-reading.
|
|
19
|
+
*/
|
|
20
|
+
const DEEPSEEK_PEAK_RULE = {
|
|
21
|
+
url: 'https://api-docs.deepseek.com/quick_start/pricing',
|
|
22
|
+
text: 'Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday, excluding Chinese public holidays.',
|
|
23
|
+
};
|
|
24
|
+
/**
|
|
25
|
+
* Every date from `from` to `to` inclusive, as YYYY-MM-DD.
|
|
26
|
+
*
|
|
27
|
+
* @param from - First date (YYYY-MM-DD).
|
|
28
|
+
* @param to - Last date (YYYY-MM-DD).
|
|
29
|
+
* @returns The dates.
|
|
30
|
+
*/
|
|
31
|
+
function dateRange(from, to) {
|
|
32
|
+
const out = [];
|
|
33
|
+
for (let t = Date.parse(`${from}T00:00:00Z`); t <= Date.parse(`${to}T00:00:00Z`); t += 86_400_000)
|
|
34
|
+
out.push(new Date(t).toISOString().slice(0, 10));
|
|
35
|
+
return out;
|
|
36
|
+
}
|
|
37
|
+
/**
|
|
38
|
+
* Chinese public holidays (the 放假 ranges), for `peakPricing.excludedDatesUtc`
|
|
39
|
+
* on the DeepSeek models: their peak is "01:00 - 04:00 and 06:00 - 10:00 UTC,
|
|
40
|
+
* Monday through Friday, excluding Chinese public holidays" (pricing page,
|
|
41
|
+
* 2026-09-23). Those windows are 09:00–12:00 and 14:00–18:00 Beijing time, so
|
|
42
|
+
* the Beijing holiday date IS the UTC date of every window.
|
|
43
|
+
*
|
|
44
|
+
* Source — the State Council's notice for 2026 (国办发明电〔2025〕7号), read
|
|
45
|
+
* 2026-09-23 at https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html.
|
|
46
|
+
* Make-up working days (调休 weekends) need no entry: the peak is Mon–Fri only.
|
|
47
|
+
*
|
|
48
|
+
* ADD NEXT YEAR'S DATES when the State Council publishes them (each November);
|
|
49
|
+
* check-model-freshness warns 45 days before a year with no dates here.
|
|
50
|
+
*/
|
|
51
|
+
const CHINA_PUBLIC_HOLIDAYS = [
|
|
52
|
+
...dateRange('2026-01-01', '2026-01-03'), // 元旦
|
|
53
|
+
...dateRange('2026-02-15', '2026-02-23'), // 春节
|
|
54
|
+
...dateRange('2026-04-04', '2026-04-06'), // 清明节
|
|
55
|
+
...dateRange('2026-05-01', '2026-05-05'), // 劳动节
|
|
56
|
+
...dateRange('2026-06-19', '2026-06-21'), // 端午节
|
|
57
|
+
...dateRange('2026-09-25', '2026-09-27'), // 中秋节
|
|
58
|
+
...dateRange('2026-10-01', '2026-10-07'), // 国庆节
|
|
59
|
+
];
|
|
15
60
|
/**
|
|
16
61
|
* All available AI models, grouped by provider, ordered from most to least capable.
|
|
17
62
|
*
|
|
@@ -58,7 +103,8 @@ const WEEKDAYS_UTC = [1, 2, 3, 4, 5];
|
|
|
58
103
|
* 2026-07-30 GPT-5.6 repricing — cross-check prices against models.dev with
|
|
59
104
|
* `npm run check:model-freshness` from the workspace root):
|
|
60
105
|
* - Anthropic: https://platform.claude.com/docs/en/about-claude/models/overview
|
|
61
|
-
* + /docs/en/build-with-claude/effort (fable-5 / opus-5 / sonnet-5
|
|
106
|
+
* + /docs/en/build-with-claude/effort (fable-5-1 / opus-5-5 / sonnet-5
|
|
107
|
+
* current as of 2026-09-23 — see the dated notes below; historically fable-5 / opus-5 / sonnet-5 current;
|
|
62
108
|
* opus-4-8 superseded by opus-5 at identical pricing but still served — it is
|
|
63
109
|
* the recommended refusal-fallback model; effort ladder on all three current
|
|
64
110
|
* 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];
|
|
|
88
134
|
* any/tool now 400s, thinking blocks are model-bound, editing earlier turns
|
|
89
135
|
* invalidates them) — Synthase sends tool_choice auto and is append-only,
|
|
90
136
|
* and the pre-commit dispatch probe covers the entry as sent.)
|
|
137
|
+
* (verified 2026-09-23 — ADDED claude-opus-5-5, released 2026-09-22 and
|
|
138
|
+
* listed as "Active (latest)" on /docs/en/models/opus-5-5/overview; the
|
|
139
|
+
* overview page now lists claude-opus-5 under "Legacy models (still
|
|
140
|
+
* available)" → superseded, with opus-4-8/4-7/4-6 repointed one hop.
|
|
141
|
+
* Pricing page: $4/$20, 5m cache write $5, cache hits $0.20 (footnote 2:
|
|
142
|
+
* 0.05× input on Opus 5.5 only), fast mode $8/$40. 1M ctx / 128K out, text +
|
|
143
|
+
* image input, reliable knowledge cutoff Jun 2026. Effort page: all five
|
|
144
|
+
* levels, default MEDIUM. What's-new page: thinking always on
|
|
145
|
+
* (disabled/budget_tokens 400), tool_choice any/tool 400, computer_20251124
|
|
146
|
+
* 400 — Synthase sends none of those. Every other Anthropic price on the
|
|
147
|
+
* pricing page is unchanged.)
|
|
91
148
|
* - OpenAI: https://developers.openai.com/api/docs/pricing (GPT-5.6 family GA
|
|
92
149
|
* 2026-07-09; REPRICED 2026-07-30: -luna cut 80% to $0.20/$1.20, -terra cut
|
|
93
150
|
* 20% to $2/$12, -sol unchanged $5/$30; cache read 0.1× input; gpt-5.5/
|
|
@@ -112,6 +169,14 @@ const WEEKDAYS_UTC = [1, 2, 3, 4, 5];
|
|
|
112
169
|
* in the catalog because it cannot serve a tool-carrying request on the
|
|
113
170
|
* bond's /v1/chat/completions endpoint; see the DO NOT ADD block above the
|
|
114
171
|
* OpenAI entries for the live 400s and the condition that lifts it.)
|
|
172
|
+
* (re-verified 2026-09-23: ADDED gpt-6-sol ($2/$10, cached $0.20, cache
|
|
173
|
+
* writes $2.50) and gpt-6-luna ($0.10/$0.50, cached $0.01, cache writes
|
|
174
|
+
* $0.125), released 2026-09-22 — both with the same unmodeled >272K band,
|
|
175
|
+
* 1.05M ctx / 128K out, image input, knowledge cutoff Apr 20 / May 18 2026
|
|
176
|
+
* per their model pages. Unlike astra they serve tool-carrying requests on
|
|
177
|
+
* /v1/chat/completions with reasoning_effort 'none'. The 5.6 family is now
|
|
178
|
+
* superseded; every 5.6 price on the page is unchanged and the -sol promo
|
|
179
|
+
* footnote still reads "at least through November 21, 2026".)
|
|
115
180
|
* - Google: https://ai.google.dev/gemini-api/docs/pricing (gemini-3.6-flash GA
|
|
116
181
|
* 2026-07-21 $1.50/$7.50 supersedes 3.5-flash as the agentic flagship;
|
|
117
182
|
* gemini-3.1-pro-preview still the pro tier — "3.5 Pro" has NOT shipped as
|
|
@@ -144,8 +209,16 @@ const WEEKDAYS_UTC = [1, 2, 3, 4, 5];
|
|
|
144
209
|
* not modeled; reasoning_effort low|medium|high default high, image input;
|
|
145
210
|
* grok-4.3 still served at $1.25/$2.50 with the bigger 1M window;
|
|
146
211
|
* grok-code-fast-1 no longer listed — retires 2026-08-15)
|
|
212
|
+
* - Chinese public holidays (DeepSeek's peak excludes them):
|
|
213
|
+
* https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html
|
|
214
|
+
* (the State Council's 2026 notice; see CHINA_PUBLIC_HOLIDAYS)
|
|
147
215
|
* - DeepSeek: https://api-docs.deepseek.com/quick_start/pricing +
|
|
148
|
-
* /updates/ (verified 2026-09-
|
|
216
|
+
* /updates/ (verified 2026-09-23 — the announced 2026-09-14 routing of
|
|
217
|
+
* `deepseek-v4-pro` to V4.1 Flash was WITHDRAWN: V4 Pro keeps its card
|
|
218
|
+
* 0.66/1.98/0.022 off-peak, so its staged `scheduledPricing` was deleted.
|
|
219
|
+
* The card now also says peak hours exclude Chinese public holidays, which
|
|
220
|
+
* `peakPricing` cannot express — peak is billed on those weekdays too.
|
|
221
|
+
* Earlier, 2026-09-10: legacy deepseek-chat/-reasoner ids fully
|
|
149
222
|
* retired 2026-07-24 — never in this catalog. The 2026-09-10 re-read caught
|
|
150
223
|
* the V4.1-Flash release DAY-OF: new evergreen id `deepseek-flash` at
|
|
151
224
|
* off-peak miss $0.15 / hit $0.003 / out $0.6 (peak ×2, same Mon-Fri UTC
|
|
@@ -230,6 +303,14 @@ const WEEKDAYS_UTC = [1, 2, 3, 4, 5];
|
|
|
230
303
|
* card again — $0.15/$0.50, cached input $0.03, no promo footnote. models.dev
|
|
231
304
|
* still carries the expired promo rate, so the KNOWN_DIVERGENCES entry now
|
|
232
305
|
* runs to a fixed 2026-12-09 re-verify date.)
|
|
306
|
+
* (verified 2026-09-23: glm-5.3-flashx ADDED — the ~200 tokens/s tier of the
|
|
307
|
+
* GLM-5.3-Flash series. Pricing page: $0.37/$1.25, cached input $0.075,
|
|
308
|
+
* cache storage free. Chat-completion reference: id `glm-5.3-flashx` in the
|
|
309
|
+
* vision request enum, 128K max output, reasoning_effort low|high|max
|
|
310
|
+
* (default max), function tools supported. GLM-5.3-Flash guide: 1M ctx,
|
|
311
|
+
* video/image/text/file input. No published knowledge cutoff. Not on
|
|
312
|
+
* DeepInfra (zai-org/GLM-5.3-FlashX 404s), so cn-region-only. glm-5.3 and
|
|
313
|
+
* glm-5.3-flash rates unchanged on the same page.)
|
|
233
314
|
*
|
|
234
315
|
* Knowledge-cutoff dates on non-Anthropic entries are best-effort estimates
|
|
235
316
|
* where the provider doesn't publish one; the provider sources above verify
|
|
@@ -247,6 +328,9 @@ export const MODELS = [
|
|
|
247
328
|
{
|
|
248
329
|
id: 'claude-fable-5-1',
|
|
249
330
|
provider: 'anthropic',
|
|
331
|
+
// temperature → 400 "`temperature` is deprecated for this model" (probed
|
|
332
|
+
// live 2026-09-23); callers omit it (request-shape.ts temperatureParam).
|
|
333
|
+
rejectsTemperature: true,
|
|
250
334
|
label: 'Claude Fable 5.1',
|
|
251
335
|
description: 'Most capable Anthropic — frontier reasoning & long-horizon agents',
|
|
252
336
|
contextWindow: 1_000_000,
|
|
@@ -260,6 +344,10 @@ export const MODELS = [
|
|
|
260
344
|
// all five levels are supported (effort docs, verified 2026-09-01).
|
|
261
345
|
// Default/recommended is high; xhigh/max for the most capability-sensitive
|
|
262
346
|
// agentic work, medium/low for routine work.
|
|
347
|
+
// tool_choice any/tool → 400 "tool_choice: type "tool" and "any" are not
|
|
348
|
+
// supported for this model" (probed live 2026-09-23). Discovery and
|
|
349
|
+
// starting-point selection send `auto` for it instead (request-shape.ts).
|
|
350
|
+
rejectsForcedToolChoice: true,
|
|
263
351
|
supportsVision: true,
|
|
264
352
|
supportsPromptCaching: true,
|
|
265
353
|
supportsTools: true,
|
|
@@ -282,6 +370,9 @@ export const MODELS = [
|
|
|
282
370
|
{
|
|
283
371
|
id: 'claude-fable-5',
|
|
284
372
|
provider: 'anthropic',
|
|
373
|
+
// temperature → 400 "`temperature` is deprecated for this model" (probed
|
|
374
|
+
// live 2026-09-23); callers omit it (request-shape.ts temperatureParam).
|
|
375
|
+
rejectsTemperature: true,
|
|
285
376
|
label: 'Claude Fable 5',
|
|
286
377
|
description: 'Most capable Anthropic — frontier reasoning & long-horizon agents',
|
|
287
378
|
contextWindow: 1_000_000,
|
|
@@ -317,9 +408,63 @@ export const MODELS = [
|
|
|
317
408
|
deprecatedAt: '2026-09-01',
|
|
318
409
|
supersededBy: 'claude-fable-5-1',
|
|
319
410
|
},
|
|
411
|
+
{
|
|
412
|
+
id: 'claude-opus-5-5',
|
|
413
|
+
provider: 'anthropic',
|
|
414
|
+
// temperature → 400 "`temperature` is deprecated for this model" (probed
|
|
415
|
+
// live 2026-09-23); callers omit it (request-shape.ts temperatureParam).
|
|
416
|
+
rejectsTemperature: true,
|
|
417
|
+
label: 'Claude Opus 5.5',
|
|
418
|
+
description: 'Anthropic Opus flagship — long-running agentic coding, cheaper than Opus 5',
|
|
419
|
+
// tool_choice any/tool → 400 "tool_choice: type "tool" and "any" are not
|
|
420
|
+
// supported for this model" (probed live 2026-09-23). Discovery and
|
|
421
|
+
// starting-point selection send `auto` for it instead (request-shape.ts).
|
|
422
|
+
rejectsForcedToolChoice: true,
|
|
423
|
+
// Fast mode (research preview, Claude API only): $8/$40 per MTok (pricing
|
|
424
|
+
// page, fast-mode table, verified 2026-09-23). Cache multipliers stack on
|
|
425
|
+
// the fast input rate: read 0.05× (this model's rate), write 1.25×.
|
|
426
|
+
fastPricing: {
|
|
427
|
+
inputPricePerMTok: 8,
|
|
428
|
+
outputPricePerMTok: 40,
|
|
429
|
+
cacheReadPricePerMTok: 0.4,
|
|
430
|
+
cacheWritePricePerMTok: 10,
|
|
431
|
+
},
|
|
432
|
+
contextWindow: 1_000_000,
|
|
433
|
+
maxOutputTokens: 128_000,
|
|
434
|
+
supportsThinking: true,
|
|
435
|
+
thinkingBudgetTokens: 16_000,
|
|
436
|
+
thinkingConfigurable: true,
|
|
437
|
+
supportedEffortLevels: ['low', 'medium', 'high', 'xhigh', 'max'],
|
|
438
|
+
defaultEffortLevel: 'medium',
|
|
439
|
+
// Adaptive thinking is ALWAYS ON: thinking {type:"disabled"} and manual
|
|
440
|
+
// budget_tokens both 400 — effort is the only depth control. All five
|
|
441
|
+
// levels supported; the API default is MEDIUM (not high like opus-5).
|
|
442
|
+
// tool_choice any/tool 400 on this model (auto/none only), as on
|
|
443
|
+
// fable-5-1. 512-token prompt-cache minimum. (Model page + what's-new,
|
|
444
|
+
// verified 2026-09-23.)
|
|
445
|
+
supportsVision: true,
|
|
446
|
+
supportsPromptCaching: true,
|
|
447
|
+
supportsTools: true,
|
|
448
|
+
webSearchToolType: 'web_search_20260209',
|
|
449
|
+
// Same server-tool versions as the rest of the 4.6+ Anthropic fleet. Not
|
|
450
|
+
// currently sent by Synthase beyond webSearchToolType.
|
|
451
|
+
codeExecutionToolType: 'code_execution_20260521',
|
|
452
|
+
webFetchToolType: 'web_fetch_20260209',
|
|
453
|
+
inputPricePerMTok: 4,
|
|
454
|
+
outputPricePerMTok: 20,
|
|
455
|
+
// Cache READ is a documented exception: 0.05× input ("$0.20 / MTok",
|
|
456
|
+
// pricing page footnote 2). 5-minute cache write is the usual 1.25×.
|
|
457
|
+
cacheReadPricePerMTok: 0.2,
|
|
458
|
+
cacheWritePricePerMTok: 5,
|
|
459
|
+
// Reliable knowledge cutoff Jun 2026 (model page "Specifications").
|
|
460
|
+
knowledgeCutoff: '2026-06-01',
|
|
461
|
+
},
|
|
320
462
|
{
|
|
321
463
|
id: 'claude-opus-5',
|
|
322
464
|
provider: 'anthropic',
|
|
465
|
+
// temperature → 400 "`temperature` is deprecated for this model" (probed
|
|
466
|
+
// live 2026-09-23); callers omit it (request-shape.ts temperatureParam).
|
|
467
|
+
rejectsTemperature: true,
|
|
323
468
|
label: 'Claude Opus 5',
|
|
324
469
|
description: 'Anthropic Opus flagship — step-change agentic coding at 4.8 pricing',
|
|
325
470
|
// Fast mode (research preview, Claude API only): same model at up to 2.5×
|
|
@@ -362,10 +507,18 @@ export const MODELS = [
|
|
|
362
507
|
cacheWritePricePerMTok: 6.25,
|
|
363
508
|
// Not published at verification time — best-effort estimate (≥ Opus 4.8's).
|
|
364
509
|
knowledgeCutoff: '2026-01-01',
|
|
510
|
+
// Superseded by claude-opus-5-5 (released 2026-09-22, same Opus tier,
|
|
511
|
+
// cheaper at $4/$20); Anthropic now lists opus-5 under "Legacy models
|
|
512
|
+
// (still available)". Still served and priceable, just not OFFERED.
|
|
513
|
+
deprecatedAt: '2026-09-22',
|
|
514
|
+
supersededBy: 'claude-opus-5-5',
|
|
365
515
|
},
|
|
366
516
|
{
|
|
367
517
|
id: 'claude-opus-4-8',
|
|
368
518
|
provider: 'anthropic',
|
|
519
|
+
// temperature → 400 "`temperature` is deprecated for this model" (probed
|
|
520
|
+
// live 2026-09-23); callers omit it (request-shape.ts temperatureParam).
|
|
521
|
+
rejectsTemperature: true,
|
|
369
522
|
label: 'Claude Opus 4.8',
|
|
370
523
|
description: 'Previous Opus — deep reasoning; the opus-5 refusal fallback',
|
|
371
524
|
contextWindow: 1_000_000,
|
|
@@ -396,12 +549,16 @@ export const MODELS = [
|
|
|
396
549
|
// Superseded by claude-opus-5 (same price, same tier); still served upstream
|
|
397
550
|
// and the recommended refusal-fallback target, so it stays priceable and
|
|
398
551
|
// callable by id — it is just not OFFERED, since opus-5 is a drop-in.
|
|
552
|
+
// `supersededBy` names the CURRENT selectable Opus (opus-5-5, one hop).
|
|
399
553
|
deprecatedAt: '2026-07-28',
|
|
400
|
-
supersededBy: 'claude-opus-5',
|
|
554
|
+
supersededBy: 'claude-opus-5-5',
|
|
401
555
|
},
|
|
402
556
|
{
|
|
403
557
|
id: 'claude-sonnet-5',
|
|
404
558
|
provider: 'anthropic',
|
|
559
|
+
// temperature → 400 "`temperature` is deprecated for this model" (probed
|
|
560
|
+
// live 2026-09-23); callers omit it (request-shape.ts temperatureParam).
|
|
561
|
+
rejectsTemperature: true,
|
|
405
562
|
label: 'Claude Sonnet 5',
|
|
406
563
|
description: 'Fast & capable — near-Opus coding at Sonnet cost',
|
|
407
564
|
contextWindow: 1_000_000,
|
|
@@ -437,6 +594,9 @@ export const MODELS = [
|
|
|
437
594
|
{
|
|
438
595
|
id: 'claude-opus-4-7',
|
|
439
596
|
provider: 'anthropic',
|
|
597
|
+
// temperature → 400 "`temperature` is deprecated for this model" (probed
|
|
598
|
+
// live 2026-09-23); callers omit it (request-shape.ts temperatureParam).
|
|
599
|
+
rejectsTemperature: true,
|
|
440
600
|
label: 'Claude Opus 4.7',
|
|
441
601
|
description: 'Older Opus — long-horizon agentic work, knowledge work & vision',
|
|
442
602
|
contextWindow: 1_000_000,
|
|
@@ -470,10 +630,10 @@ export const MODELS = [
|
|
|
470
630
|
// still Active upstream (deprecations page 2026-08-06: retires no sooner
|
|
471
631
|
// than 2027-04-16). NO fast mode — speed:"fast" on 4.7 returns an error
|
|
472
632
|
// (pricing page, fast-mode section). `supersededBy` names the CURRENT
|
|
473
|
-
// selectable Opus (opus-5), not the also-superseded 4.8, so a saved
|
|
633
|
+
// selectable Opus (opus-5-5), not the also-superseded 4.8 / 5, so a saved
|
|
474
634
|
// selection resolves forward in one hop.
|
|
475
635
|
deprecatedAt: '2026-05-28',
|
|
476
|
-
supersededBy: 'claude-opus-5',
|
|
636
|
+
supersededBy: 'claude-opus-5-5',
|
|
477
637
|
},
|
|
478
638
|
{
|
|
479
639
|
id: 'claude-opus-4-6',
|
|
@@ -505,9 +665,9 @@ export const MODELS = [
|
|
|
505
665
|
cacheReadPricePerMTok: 0.5,
|
|
506
666
|
cacheWritePricePerMTok: 6.25,
|
|
507
667
|
knowledgeCutoff: '2025-05-01',
|
|
508
|
-
// Superseded by the current Opus (opus-5) — kept priceable, not offered.
|
|
668
|
+
// Superseded by the current Opus (opus-5-5) — kept priceable, not offered.
|
|
509
669
|
deprecatedAt: '2026-06-16',
|
|
510
|
-
supersededBy: 'claude-opus-5',
|
|
670
|
+
supersededBy: 'claude-opus-5-5',
|
|
511
671
|
},
|
|
512
672
|
{
|
|
513
673
|
id: 'claude-sonnet-4-6',
|
|
@@ -612,7 +772,74 @@ export const MODELS = [
|
|
|
612
772
|
// Note also that the docs page advertises effort 'max', which
|
|
613
773
|
// /v1/chat/completions rejects for this model — verify the ladder against the
|
|
614
774
|
// endpoint, not the docs, when this is revisited.
|
|
775
|
+
//
|
|
776
|
+
// gpt-6-sol and gpt-6-luna (released 2026-09-22) are NOT astra's case: both
|
|
777
|
+
// still accept reasoning_effort 'none', and tools + 'none' returns 200 on
|
|
778
|
+
// /v1/chat/completions (probed live 2026-09-23 on our own key), so the same
|
|
779
|
+
// `toolsRequireReasoningOff` pin that serves the 5.6 family serves them. On
|
|
780
|
+
// that endpoint both accept none|low|medium|high|xhigh and reject 'max' —
|
|
781
|
+
// again despite the docs pages listing it. GPT-6 has no Terra tier: sol at
|
|
782
|
+
// $2/$10 is priced at 5.6-terra's tier and undercuts 5.6-sol, so it supersedes
|
|
783
|
+
// both; luna supersedes 5.6-luna at half the price.
|
|
615
784
|
// ---------------------------------------------------------------------------
|
|
785
|
+
{
|
|
786
|
+
id: 'gpt-6-sol',
|
|
787
|
+
provider: 'openai',
|
|
788
|
+
label: 'GPT-6 Sol',
|
|
789
|
+
description: 'OpenAI for complex coding & agentic work',
|
|
790
|
+
// Documented as 1.05M; floored to 1M like the 5.6 entries.
|
|
791
|
+
contextWindow: 1_000_000,
|
|
792
|
+
maxOutputTokens: 128_000,
|
|
793
|
+
supportsThinking: true,
|
|
794
|
+
thinkingBudgetTokens: 16_000,
|
|
795
|
+
thinkingConfigurable: true,
|
|
796
|
+
supportedEffortLevels: ['low', 'medium', 'high', 'xhigh'],
|
|
797
|
+
defaultEffortLevel: 'medium',
|
|
798
|
+
supportsVision: true,
|
|
799
|
+
supportsPromptCaching: true,
|
|
800
|
+
supportsTools: true,
|
|
801
|
+
// Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
|
|
802
|
+
toolsRequireReasoningOff: true,
|
|
803
|
+
// NO webSearchToolType: see gpt-5.6-sol — the bond calls
|
|
804
|
+
// /v1/chat/completions, which has no web_search tool type. Re-add when the
|
|
805
|
+
// bond moves to /v1/responses.
|
|
806
|
+
codeExecutionToolType: 'code_interpreter',
|
|
807
|
+
// Standard tier. A long-context band above 272K prompt tokens reprices the
|
|
808
|
+
// whole request ($4/$15, cached $0.40, cache writes $5) — not modeled, same
|
|
809
|
+
// as the other >200K tiers.
|
|
810
|
+
inputPricePerMTok: 2,
|
|
811
|
+
outputPricePerMTok: 10,
|
|
812
|
+
cacheReadPricePerMTok: 0.2,
|
|
813
|
+
cacheWritePricePerMTok: 2.5,
|
|
814
|
+
knowledgeCutoff: '2026-04-20',
|
|
815
|
+
},
|
|
816
|
+
{
|
|
817
|
+
id: 'gpt-6-luna',
|
|
818
|
+
provider: 'openai',
|
|
819
|
+
label: 'GPT-6 Luna',
|
|
820
|
+
description: 'Fast & cheap OpenAI — light tasks & subagents',
|
|
821
|
+
contextWindow: 1_000_000,
|
|
822
|
+
maxOutputTokens: 128_000,
|
|
823
|
+
supportsThinking: true,
|
|
824
|
+
thinkingBudgetTokens: 8_000,
|
|
825
|
+
thinkingConfigurable: true,
|
|
826
|
+
supportedEffortLevels: ['low', 'medium', 'high', 'xhigh'],
|
|
827
|
+
defaultEffortLevel: 'medium',
|
|
828
|
+
supportsVision: true,
|
|
829
|
+
supportsPromptCaching: true,
|
|
830
|
+
supportsTools: true,
|
|
831
|
+
// Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
|
|
832
|
+
toolsRequireReasoningOff: true,
|
|
833
|
+
// NO webSearchToolType: see gpt-5.6-sol.
|
|
834
|
+
codeExecutionToolType: 'code_interpreter',
|
|
835
|
+
// Standard tier; >272K band ($0.20/$0.75, cached $0.02, writes $0.25) not
|
|
836
|
+
// modeled.
|
|
837
|
+
inputPricePerMTok: 0.1,
|
|
838
|
+
outputPricePerMTok: 0.5,
|
|
839
|
+
cacheReadPricePerMTok: 0.01,
|
|
840
|
+
cacheWritePricePerMTok: 0.125,
|
|
841
|
+
knowledgeCutoff: '2026-05-18',
|
|
842
|
+
},
|
|
616
843
|
{
|
|
617
844
|
id: 'gpt-5.6-sol',
|
|
618
845
|
provider: 'openai',
|
|
@@ -651,6 +878,10 @@ export const MODELS = [
|
|
|
651
878
|
cacheWritePricePerMTok: 6.25,
|
|
652
879
|
// Not published — best-effort estimate.
|
|
653
880
|
knowledgeCutoff: '2026-03-01',
|
|
881
|
+
// Superseded by gpt-6-sol ($2/$10 vs $5/$30 list). Still served and
|
|
882
|
+
// priceable — just not offered in the picker.
|
|
883
|
+
deprecatedAt: '2026-09-23',
|
|
884
|
+
supersededBy: 'gpt-6-sol',
|
|
654
885
|
},
|
|
655
886
|
{
|
|
656
887
|
id: 'gpt-5.6-terra',
|
|
@@ -683,6 +914,10 @@ export const MODELS = [
|
|
|
683
914
|
cacheWritePricePerMTok: 2.5,
|
|
684
915
|
// Not published — best-effort estimate.
|
|
685
916
|
knowledgeCutoff: '2026-03-01',
|
|
917
|
+
// Superseded by gpt-6-sol: GPT-6 has no Terra tier, and sol sits at this
|
|
918
|
+
// tier's price ($2/$10 vs $2/$12).
|
|
919
|
+
deprecatedAt: '2026-09-23',
|
|
920
|
+
supersededBy: 'gpt-6-sol',
|
|
686
921
|
},
|
|
687
922
|
{
|
|
688
923
|
id: 'gpt-5.6-luna',
|
|
@@ -720,6 +955,9 @@ export const MODELS = [
|
|
|
720
955
|
// note for the same removal pattern).
|
|
721
956
|
// Not published — best-effort estimate.
|
|
722
957
|
knowledgeCutoff: '2026-03-01',
|
|
958
|
+
// Superseded by gpt-6-luna ($0.10/$0.50 — half the price).
|
|
959
|
+
deprecatedAt: '2026-09-23',
|
|
960
|
+
supersededBy: 'gpt-6-luna',
|
|
723
961
|
},
|
|
724
962
|
{
|
|
725
963
|
id: 'gpt-5.5',
|
|
@@ -752,10 +990,11 @@ export const MODELS = [
|
|
|
752
990
|
cacheReadPricePerMTok: 0.5,
|
|
753
991
|
cacheWritePricePerMTok: 5,
|
|
754
992
|
knowledgeCutoff: '2025-12-01',
|
|
755
|
-
// Superseded by gpt-5.6-sol (same frontier tier, same $5/$30)
|
|
756
|
-
//
|
|
993
|
+
// Superseded by gpt-5.6-sol (same frontier tier, same $5/$30), and since
|
|
994
|
+
// 2026-09-23 points one hop to gpt-6-sol, 5.6-sol's own successor. Still
|
|
995
|
+
// listed as current by OpenAI, so it stays priceable — just not offered.
|
|
757
996
|
deprecatedAt: '2026-07-09',
|
|
758
|
-
supersededBy: 'gpt-
|
|
997
|
+
supersededBy: 'gpt-6-sol',
|
|
759
998
|
},
|
|
760
999
|
{
|
|
761
1000
|
id: 'gpt-5.4',
|
|
@@ -789,9 +1028,10 @@ export const MODELS = [
|
|
|
789
1028
|
// OpenAI still lists gpt-5.4 as current, but gpt-5.6-terra covers this
|
|
790
1029
|
// balanced tier for LESS ($2/$12 vs $2.50/$15) — superseded, so the picker
|
|
791
1030
|
// offers only the 5.6 generation (this is OUR taxonomy, not OpenAI's
|
|
792
|
-
// deprecations page; the model stays priceable).
|
|
1031
|
+
// deprecations page; the model stays priceable). Points one hop to
|
|
1032
|
+
// gpt-6-sol since 2026-09-23, when 5.6-terra was itself superseded.
|
|
793
1033
|
deprecatedAt: '2026-07-28',
|
|
794
|
-
supersededBy: 'gpt-
|
|
1034
|
+
supersededBy: 'gpt-6-sol',
|
|
795
1035
|
},
|
|
796
1036
|
{
|
|
797
1037
|
id: 'gpt-5.4-mini',
|
|
@@ -825,9 +1065,10 @@ export const MODELS = [
|
|
|
825
1065
|
// Superseded by gpt-5.6-luna, which IS the newer cheap/fast tier and is
|
|
826
1066
|
// strictly better on every axis that made this the budget pick: $0.20/$1.20
|
|
827
1067
|
// vs $0.75/$4.50 after the 2026-07-30 repricing, and a 1M window vs 400K.
|
|
828
|
-
// Hiding it therefore costs OpenAI no cheap option. Stays priceable.
|
|
1068
|
+
// Hiding it therefore costs OpenAI no cheap option. Stays priceable. Points
|
|
1069
|
+
// one hop to gpt-6-luna since 2026-09-23, when 5.6-luna was superseded.
|
|
829
1070
|
deprecatedAt: '2026-08-01',
|
|
830
|
-
supersededBy: 'gpt-
|
|
1071
|
+
supersededBy: 'gpt-6-luna',
|
|
831
1072
|
},
|
|
832
1073
|
// ---------------------------------------------------------------------------
|
|
833
1074
|
// Google
|
|
@@ -1213,8 +1454,10 @@ export const MODELS = [
|
|
|
1213
1454
|
// leans on that "temporary" routing until molecule-dev's default ids
|
|
1214
1455
|
// move to `deepseek-flash`.
|
|
1215
1456
|
// 2. From 2026-09-14 12:00 Beijing (04:00 UTC), until V4.1 Pro ships, ALL
|
|
1216
|
-
// `deepseek-v4-pro` requests route to V4.1 Flash at Flash
|
|
1217
|
-
//
|
|
1457
|
+
// `deepseek-v4-pro` requests were to route to V4.1 Flash at Flash
|
|
1458
|
+
// prices. WITHDRAWN before it landed (re-read 2026-09-23): V4 Pro keeps
|
|
1459
|
+
// being served after 2026-09-14 "with the billing method remaining
|
|
1460
|
+
// unchanged", so the staged `scheduledPricing` was deleted.
|
|
1218
1461
|
// 2026-08-13: V4-Pro GA — and with it the price rise that the "coming soon"
|
|
1219
1462
|
// note below had been waiting on. It was staged as `scheduledPricing`
|
|
1220
1463
|
// effective 2026-08-16T16:00Z; that instant has PASSED and the rates are now
|
|
@@ -1315,33 +1558,25 @@ export const MODELS = [
|
|
|
1315
1558
|
{ startMinuteUtc: 360, endMinuteUtc: 600, daysOfWeekUtc: WEEKDAYS_UTC },
|
|
1316
1559
|
],
|
|
1317
1560
|
multiplier: 2,
|
|
1561
|
+
excludedDatesUtc: CHINA_PUBLIC_HOLIDAYS,
|
|
1562
|
+
rule: DEEPSEEK_PEAK_RULE,
|
|
1318
1563
|
},
|
|
1319
|
-
//
|
|
1320
|
-
//
|
|
1321
|
-
//
|
|
1322
|
-
//
|
|
1323
|
-
//
|
|
1324
|
-
//
|
|
1325
|
-
//
|
|
1326
|
-
//
|
|
1327
|
-
|
|
1328
|
-
effectiveFrom: '2026-09-14T04:00:00Z',
|
|
1329
|
-
inputPricePerMTok: 0.15,
|
|
1330
|
-
outputPricePerMTok: 0.6,
|
|
1331
|
-
cacheReadPricePerMTok: 0.003,
|
|
1332
|
-
cacheWritePricePerMTok: 0.15,
|
|
1333
|
-
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',
|
|
1334
|
-
},
|
|
1564
|
+
// The 2026-09-10 announcement that this id would route to V4.1 Flash at
|
|
1565
|
+
// Flash prices from 2026-09-14 was WITHDRAWN: the updates page now reads
|
|
1566
|
+
// "we have decided to continue providing API services for DeepSeek V4 Pro
|
|
1567
|
+
// after September 14, 2026, with the billing method remaining unchanged",
|
|
1568
|
+
// and the pricing page still lists `deepseek-v4-pro` (V4-Pro-0813) at
|
|
1569
|
+
// 0.66/1.98, cache hit 0.022 off-peak (both re-read 2026-09-23). So the
|
|
1570
|
+
// staged `scheduledPricing` flash card was deleted rather than folded in —
|
|
1571
|
+
// the base rates above are what DeepSeek bills. The US `regionPricing` is
|
|
1572
|
+
// DeepInfra's own card and was never affected.
|
|
1335
1573
|
// Not published by DeepSeek — best-effort estimate.
|
|
1336
1574
|
knowledgeCutoff: '2025-07-01',
|
|
1337
|
-
//
|
|
1338
|
-
// V4.1 Flash
|
|
1339
|
-
//
|
|
1340
|
-
//
|
|
1341
|
-
//
|
|
1342
|
-
// this one's. Until the 14th it still serves real V4-Pro-0813 weights at
|
|
1343
|
-
// the Pro card, so it stays selectable for existing selections.
|
|
1344
|
-
deprecatedAt: '2026-09-14',
|
|
1575
|
+
// Was deprecated 2026-09-14 on the expectation that the id would stop
|
|
1576
|
+
// being itself (routed to V4.1 Flash). DeepSeek withdrew that routing — V4
|
|
1577
|
+
// Pro continues "with the billing method remaining unchanged" (updates
|
|
1578
|
+
// page, re-read 2026-09-23) — so it is offered again (owner, 2026-09-23:
|
|
1579
|
+
// "keep offering deepseek pro if it is available").
|
|
1345
1580
|
},
|
|
1346
1581
|
{
|
|
1347
1582
|
id: 'deepseek-v4-flash',
|
|
@@ -1424,6 +1659,8 @@ export const MODELS = [
|
|
|
1424
1659
|
{ startMinuteUtc: 360, endMinuteUtc: 600, daysOfWeekUtc: WEEKDAYS_UTC },
|
|
1425
1660
|
],
|
|
1426
1661
|
multiplier: 2,
|
|
1662
|
+
excludedDatesUtc: CHINA_PUBLIC_HOLIDAYS,
|
|
1663
|
+
rule: DEEPSEEK_PEAK_RULE,
|
|
1427
1664
|
},
|
|
1428
1665
|
// Not published by DeepSeek — best-effort estimate.
|
|
1429
1666
|
knowledgeCutoff: '2025-07-01',
|
|
@@ -1488,6 +1725,8 @@ export const MODELS = [
|
|
|
1488
1725
|
{ startMinuteUtc: 360, endMinuteUtc: 600, daysOfWeekUtc: WEEKDAYS_UTC },
|
|
1489
1726
|
],
|
|
1490
1727
|
multiplier: 2,
|
|
1728
|
+
excludedDatesUtc: CHINA_PUBLIC_HOLIDAYS,
|
|
1729
|
+
rule: DEEPSEEK_PEAK_RULE,
|
|
1491
1730
|
},
|
|
1492
1731
|
// Not published by DeepSeek — best-effort estimate (family estimate; the
|
|
1493
1732
|
// V4.1 announcement lists benchmarks but no training cutoff).
|
|
@@ -1514,6 +1753,9 @@ export const MODELS = [
|
|
|
1514
1753
|
{
|
|
1515
1754
|
id: 'kimi-k3',
|
|
1516
1755
|
provider: 'moonshot',
|
|
1756
|
+
// Native Moonshot host: temperature 0 → 400 "invalid temperature: only 1 is
|
|
1757
|
+
// allowed for this model" (probed 2026-09-23); callers omit it.
|
|
1758
|
+
rejectsTemperature: true,
|
|
1517
1759
|
label: 'Kimi K3',
|
|
1518
1760
|
description: 'Moonshot flagship — 2.8T open weights, 1M context, multimodal',
|
|
1519
1761
|
contextWindow: 1_000_000,
|
|
@@ -1557,6 +1799,12 @@ export const MODELS = [
|
|
|
1557
1799
|
{
|
|
1558
1800
|
id: 'kimi-k2.7-code',
|
|
1559
1801
|
provider: 'moonshot',
|
|
1802
|
+
// Native Moonshot host (probed 2026-09-23): temperature 0 → 400 "only 1 is
|
|
1803
|
+
// allowed"; and tool_choice 'required' → 400 "incompatible with thinking
|
|
1804
|
+
// enabled" — thinking cannot be turned off on this model, so a forced tool
|
|
1805
|
+
// call goes out as `auto` (request-shape.ts).
|
|
1806
|
+
rejectsTemperature: true,
|
|
1807
|
+
rejectsForcedToolChoice: true,
|
|
1560
1808
|
label: 'Kimi K2.7 Code',
|
|
1561
1809
|
description: 'Moonshot coding specialist — token-efficient agentic coding',
|
|
1562
1810
|
contextWindow: 262_144,
|
|
@@ -2003,6 +2251,43 @@ export const MODELS = [
|
|
|
2003
2251
|
// publishes no cutoff.
|
|
2004
2252
|
knowledgeCutoff: '2025-06-01',
|
|
2005
2253
|
},
|
|
2254
|
+
{
|
|
2255
|
+
// The high-speed serving tier of the GLM-5.3-Flash series (~200 tokens/s
|
|
2256
|
+
// per the GLM-5.3-Flash guide) — same series surface, higher rate card.
|
|
2257
|
+
// Same 5.3 generation as glm-5.3-flash, so neither supersedes the other.
|
|
2258
|
+
id: 'glm-5.3-flashx',
|
|
2259
|
+
provider: 'zhipu',
|
|
2260
|
+
label: 'GLM-5.3 FlashX',
|
|
2261
|
+
description: 'High-speed native multimodal coding + agents — 1M context',
|
|
2262
|
+
contextWindow: 1_048_576,
|
|
2263
|
+
// "GLM-5.3-Flash series supports a maximum output length of 128K" —
|
|
2264
|
+
// chat-completion API reference, checked 2026-09-23.
|
|
2265
|
+
maxOutputTokens: 131_072,
|
|
2266
|
+
supportsThinking: true,
|
|
2267
|
+
thinkingBudgetTokens: 8_000,
|
|
2268
|
+
thinkingConfigurable: true,
|
|
2269
|
+
// Same surface as glm-5.3-flash: low|high|max only, default max, thinking
|
|
2270
|
+
// cannot be disabled. high is the balanced tier we default to.
|
|
2271
|
+
supportedEffortLevels: ['low', 'high', 'max'],
|
|
2272
|
+
defaultEffortLevel: 'high',
|
|
2273
|
+
// Series input: text, images, video, files (glm-5.3-flashx appears in the
|
|
2274
|
+
// chat-completion reference's image examples).
|
|
2275
|
+
supportsVision: true,
|
|
2276
|
+
supportsPromptCaching: true,
|
|
2277
|
+
supportsTools: true,
|
|
2278
|
+
webSearchToolType: 'web_search',
|
|
2279
|
+
// List card, verified 2026-09-23 on the Z.ai pricing page.
|
|
2280
|
+
inputPricePerMTok: 0.37,
|
|
2281
|
+
outputPricePerMTok: 1.25,
|
|
2282
|
+
// GLM context cache: read ≈0.2× input, no write premium (storage free).
|
|
2283
|
+
cacheReadPricePerMTok: 0.075,
|
|
2284
|
+
cacheWritePricePerMTok: 0.37,
|
|
2285
|
+
// Native host only: api.deepinfra.com/models/zai-org/GLM-5.3-FlashX is a
|
|
2286
|
+
// 404 (checked 2026-09-23), so there is no US re-host to map.
|
|
2287
|
+
regions: ['cn'],
|
|
2288
|
+
// Not published by Z.ai — best-effort estimate for the 5.3 generation.
|
|
2289
|
+
knowledgeCutoff: '2025-06-01',
|
|
2290
|
+
},
|
|
2006
2291
|
{
|
|
2007
2292
|
id: 'glm-5.3-flash',
|
|
2008
2293
|
provider: 'zhipu',
|
package/dist/types.d.ts
CHANGED
|
@@ -131,6 +131,31 @@ export interface ModelDefinition {
|
|
|
131
131
|
* both together; until then, working-without-reasoning beats 400.
|
|
132
132
|
*/
|
|
133
133
|
toolsRequireReasoningOff?: boolean;
|
|
134
|
+
/**
|
|
135
|
+
* The provider rejects a FORCED tool choice for this model — Anthropic
|
|
136
|
+
* `tool_choice` `any` / `tool` answer 400 on claude-fable-5-1 and
|
|
137
|
+
* claude-opus-5-5, and Moonshot rejects `required` while thinking is on,
|
|
138
|
+
* which it always is on kimi-k2.7-code (all probed live 2026-09-23). `auto`
|
|
139
|
+
* works and the model still calls the tool. Callers that would force a tool call (Synthase
|
|
140
|
+
* discovery, starting-point selection) must send `auto` for these models;
|
|
141
|
+
* molecule-dev resolves that in one place (`toolChoiceParam` in
|
|
142
|
+
* request-shape.ts) and its live dispatch check probes the forced shape for
|
|
143
|
+
* every model, so a new model with this restriction fails the check instead
|
|
144
|
+
* of every discovery turn.
|
|
145
|
+
*/
|
|
146
|
+
rejectsForcedToolChoice?: boolean;
|
|
147
|
+
/**
|
|
148
|
+
* The provider rejects a caller-chosen `temperature` for this model —
|
|
149
|
+
* Anthropic answers 400 "`temperature` is deprecated for this model" on the
|
|
150
|
+
* Claude 5 family and Opus 4.7/4.8 (Opus 4.6, Sonnet 4.6 and Haiku 4.5 still
|
|
151
|
+
* accept it), and Moonshot's native host allows only 1 on kimi-k3 and
|
|
152
|
+
* kimi-k2.7-code (all probed live 2026-09-23). Omitting it is always safe. Callers that set a temperature (commit
|
|
153
|
+
* messages, starting-point selection) must omit it for these models;
|
|
154
|
+
* molecule-dev resolves that in one place (`temperatureParam` in
|
|
155
|
+
* request-shape.ts), and its live dispatch check sends the tool-less,
|
|
156
|
+
* temperature-0 shape to every model.
|
|
157
|
+
*/
|
|
158
|
+
rejectsTemperature?: boolean;
|
|
134
159
|
/**
|
|
135
160
|
* Provider-specific server tool type for web search (e.g. `'web_search_20250305'`).
|
|
136
161
|
* When set, the chat handler sends this as a ServerTool alongside custom tools.
|
|
@@ -234,6 +259,20 @@ export interface ModelDefinition {
|
|
|
234
259
|
* over-bill this field exists to prevent. A wrapping window belongs to the
|
|
235
260
|
* day it STARTS on, so its post-midnight tail is still matched against the
|
|
236
261
|
* previous day.
|
|
262
|
+
*
|
|
263
|
+
* `excludedDatesUtc` lists `YYYY-MM-DD` dates on which no window applies —
|
|
264
|
+
* a provider's public holidays. DeepSeek's peak excludes Chinese public
|
|
265
|
+
* holidays (its pricing page, 2026-09-23); without the list those days bill
|
|
266
|
+
* at 2× while the provider charges off-peak. A date is matched against the
|
|
267
|
+
* day a window STARTS on (UTC), like `daysOfWeekUtc`. The list is DATA that
|
|
268
|
+
* runs out: check-model-freshness warns when the coming year has no dates.
|
|
269
|
+
*
|
|
270
|
+
* `rule` is the provider's own sentence defining these windows, verbatim,
|
|
271
|
+
* and the page that publishes it. check-model-freshness re-reads the page on
|
|
272
|
+
* every run and warns the moment the sentence changes — the windows are only
|
|
273
|
+
* as right as the last reading, and DeepSeek has amended its rule twice
|
|
274
|
+
* (a weekday qualifier by 2026-08-31, a holiday exclusion by 2026-09-23)
|
|
275
|
+
* without anything here noticing.
|
|
237
276
|
*/
|
|
238
277
|
peakPricing?: {
|
|
239
278
|
windows: {
|
|
@@ -242,6 +281,11 @@ export interface ModelDefinition {
|
|
|
242
281
|
daysOfWeekUtc?: number[];
|
|
243
282
|
}[];
|
|
244
283
|
multiplier: number;
|
|
284
|
+
excludedDatesUtc?: string[];
|
|
285
|
+
rule?: {
|
|
286
|
+
url: string;
|
|
287
|
+
text: string;
|
|
288
|
+
};
|
|
245
289
|
};
|
|
246
290
|
/**
|
|
247
291
|
* A price change the provider has ANNOUNCED with a dated effective instant,
|
|
@@ -294,6 +338,11 @@ export interface ModelDefinition {
|
|
|
294
338
|
daysOfWeekUtc?: number[];
|
|
295
339
|
}[];
|
|
296
340
|
multiplier: number;
|
|
341
|
+
excludedDatesUtc?: string[];
|
|
342
|
+
rule?: {
|
|
343
|
+
url: string;
|
|
344
|
+
text: string;
|
|
345
|
+
};
|
|
297
346
|
};
|
|
298
347
|
/** Where the change was announced, for the re-verify pass after it lands. */
|
|
299
348
|
source?: string;
|
package/dist/types.d.ts.map
CHANGED
|
@@ -1 +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
|
|
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;;;;;;;;;;;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
CHANGED