@molecule/api-resource-ai-models 1.20.1 → 1.22.0
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 +15 -5
- package/dist/models.d.ts +50 -4
- package/dist/models.d.ts.map +1 -1
- package/dist/models.js +252 -32
- package/package.json +4 -4
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: 2026-11-
|
|
6
|
+
Generated: 2026-11-18T16:18:51.959Z
|
|
7
7
|
-->
|
|
8
8
|
|
|
9
9
|
# @molecule/api-resource-ai-models
|
|
@@ -112,9 +112,15 @@ interface ModelDefinition {
|
|
|
112
112
|
webSearchToolType?: string
|
|
113
113
|
/** USD per 1,000 web searches the provider bills on top of tokens, applied to `TokenUsage.webSearchRequests`. */
|
|
114
114
|
webSearchPricePer1k?: number
|
|
115
|
-
/**
|
|
115
|
+
/**
|
|
116
|
+
* Provider-specific server tool type for code execution (e.g. `'code_execution_20250825'`).
|
|
117
|
+
* Omit if the model / provider does not support native code execution.
|
|
118
|
+
*/
|
|
116
119
|
codeExecutionToolType?: string
|
|
117
|
-
/**
|
|
120
|
+
/**
|
|
121
|
+
* Provider-specific server tool type for web fetch / URL context (e.g. `'web_fetch_20260209'`).
|
|
122
|
+
* Omit if the model / provider does not support native web fetch.
|
|
123
|
+
*/
|
|
118
124
|
webFetchToolType?: string
|
|
119
125
|
/** Whether this model is available on the free tier (only one model should be true). */
|
|
120
126
|
freeTier?: boolean
|
|
@@ -170,7 +176,11 @@ interface ModelDefinition {
|
|
|
170
176
|
excludedDatesUtc?: string[]
|
|
171
177
|
rule?: { url: string; text: string }
|
|
172
178
|
}
|
|
173
|
-
/**
|
|
179
|
+
/**
|
|
180
|
+
* Long-context band from `effectiveFrom` (omitted → the model's existing
|
|
181
|
+
* band carries through unchanged). Indexed off the model-level field so
|
|
182
|
+
* the two shapes can never drift apart.
|
|
183
|
+
*/
|
|
174
184
|
longContextPricing?: ModelDefinition['longContextPricing']
|
|
175
185
|
/** Where the change was announced, for the re-verify pass after it lands. */
|
|
176
186
|
source?: string
|
|
@@ -210,7 +220,7 @@ interface ModelDefinition {
|
|
|
210
220
|
}
|
|
211
221
|
```
|
|
212
222
|
|
|
213
|
-
_(
|
|
223
|
+
_(The 21 longest member comments are shortened to their first sentence to fit; the full declaration is in `node_modules/@molecule/api-resource-ai-models/dist/types.d.ts`.)_
|
|
214
224
|
|
|
215
225
|
#### `ModeModelDefaults`
|
|
216
226
|
|
package/dist/models.d.ts
CHANGED
|
@@ -186,6 +186,28 @@ import type { ModelDefinition } from './types.js';
|
|
|
186
186
|
* the Responses API" — the same shape as gpt-6.1-sol, served by the bond's
|
|
187
187
|
* /v1/responses calls. Supersedes gpt-5.6-terra, same tier at $2/$12,
|
|
188
188
|
* repointed from gpt-6.1-sol.)
|
|
189
|
+
* (verified 2026-11-18 on the pricing page + the gpt-6.2-astra model page:
|
|
190
|
+
* ADDED gpt-6.2-astra ($8/$40, cached $0.40 — 5% of input — cache writes
|
|
191
|
+
* $10; >272K band $16/$60, cached $0.80, writes $20 not modeled; 1.05M ctx /
|
|
192
|
+
* 128K out; image input; knowledge cutoff May 18 2026; effort
|
|
193
|
+
* low|medium|high|xhigh|max with NO 'none' — live-probed on /v1/responses
|
|
194
|
+
* with tools at every level 2026-11-18, and the 'none' rejection's error
|
|
195
|
+
* names exactly those five; temperature → 400; forced tool_choice → 200).
|
|
196
|
+
* It supersedes gpt-6-astra (the same flagship tier at $10/$50). The 5.6-sol
|
|
197
|
+
* promo footnote now reads "available until November 30, 2026" — the
|
|
198
|
+
* KNOWN_DIVERGENCES expiry in check-model-freshness.mjs was extended to
|
|
199
|
+
* match. Web search stays $10/1k calls on reasoning models.)
|
|
200
|
+
* (verified 2026-11-18 on the pricing page + the gpt-6.2-sol model page:
|
|
201
|
+
* ADDED gpt-6.2-sol ($2/$10, cached $0.10 — 5% of input — cache writes
|
|
202
|
+
* $2.50; >272K band $4/$15, cached $0.20, writes $5 not modeled; 1.05M ctx /
|
|
203
|
+
* 128K out; image input; knowledge cutoff Jun 1 2026; effort
|
|
204
|
+
* low|medium|high|xhigh|max with NO 'none' — the model page states the API
|
|
205
|
+
* rejects 'none', and all five levels were probed live with tools on
|
|
206
|
+
* /v1/responses 2026-11-18; temperature → 400; forced tool_choice → 200).
|
|
207
|
+
* It supersedes gpt-6.1-sol (same $2/$10 tier, newer generation); gpt-6-sol
|
|
208
|
+
* / 5.6-sol / 5.5 / 5.4 repointed one hop to it. No GPT-6.2 terra or luna id
|
|
209
|
+
* exists on either page. The 5.6-sol promo footnote ("until November 30,
|
|
210
|
+
* 2026") and web-search price are unchanged.)
|
|
189
211
|
* - Google: https://ai.google.dev/gemini-api/docs/pricing (gemini-3.6-flash GA
|
|
190
212
|
* 2026-07-21 $1.50/$7.50 supersedes 3.5-flash as the agentic flagship;
|
|
191
213
|
* gemini-3.1-pro-preview still the pro tier — "3.5 Pro" has NOT shipped as
|
|
@@ -255,11 +277,17 @@ import type { ModelDefinition } from './types.js';
|
|
|
255
277
|
* cataloged superseded by grok-4.7.
|
|
256
278
|
* grok-4.5 still served at $2/$6, cached $0.30; grok-4.3 still served at
|
|
257
279
|
* $1.25/$2.50 with the bigger 1M window; grok-code-fast-1 no longer listed)
|
|
258
|
-
* - Chinese public holidays (DeepSeek's peak excludes them):
|
|
259
|
-
*
|
|
260
|
-
*
|
|
280
|
+
* - Chinese public holidays (DeepSeek's peak excludes them): the State
|
|
281
|
+
* Council's notices — 2026 at
|
|
282
|
+
* https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html,
|
|
283
|
+
* 2027 (published 2026-11-05, read 2026-11-17) at
|
|
284
|
+
* https://paper.people.com.cn/rmrb/pc/content/202611/06/content_30189674.html
|
|
285
|
+
* (see CHINA_PUBLIC_HOLIDAYS)
|
|
261
286
|
* - DeepSeek: https://api-docs.deepseek.com/quick_start/pricing +
|
|
262
|
-
* /updates/ (
|
|
287
|
+
* /updates/ (2026-11-17: the 2027 State Council holiday notice landed above —
|
|
288
|
+
* no DeepSeek card change; this morning's freshness-gate chunk read found the
|
|
289
|
+
* peak-rule sentence and every price unchanged, so the 2026-10-20
|
|
290
|
+
* verification stands. Earlier, verified 2026-10-20 — the whole rate card RE-READ, unchanged:
|
|
263
291
|
* pro 0.66/1.98/0.022 off-peak ×2 peak, flash 0.15/0.6/0.003, peak windows
|
|
264
292
|
* and the Chinese-public-holidays exclusion verbatim as before. The page is
|
|
265
293
|
* now a client-rendered SPA — its HTML is an app shell and the card text
|
|
@@ -374,6 +402,24 @@ import type { ModelDefinition } from './types.js';
|
|
|
374
402
|
* model page still describes it (docs lag the endpoint). The catalog entry
|
|
375
403
|
* dropped its cn region and is now US/DeepInfra-only, whose rates re-verified
|
|
376
404
|
* unchanged the same day: $0.30/$1, cache read 0.3333× input.)
|
|
405
|
+
* (re-verified 2026-11-17 across the provider's own pages: qwen4-max
|
|
406
|
+
* (released 2026-11-16) ADDED. alibabacloud.com/help/en/model-studio/
|
|
407
|
+
* qwen4-max + the EN model-pricing page: Singapore intl $1.5/$4.5 with
|
|
408
|
+
* explicit/implicit cache read $0.075 (5% of input) and every Global region
|
|
409
|
+
* at $1.413/$4.24 with cache read $0.071 — the ZH context-cache doc lists
|
|
410
|
+
* qwen4-max OUTSIDE its standard 20% cache-read table and defers to the
|
|
411
|
+
* console, whose per-region figures the pricing page publishes (explicit
|
|
412
|
+
* 5m creation 125%, 1h 200%; implicit creation 1× input). 1M ctx / 131,072
|
|
413
|
+
* out (262,144 thinking), Text+Image+Video input in every region, hybrid
|
|
414
|
+
* thinking ON by default (thinking_budget, native reasoning_effort
|
|
415
|
+
* low|high|xhigh exists but the bond sends the budget form), parallel tool
|
|
416
|
+
* calling. US = DeepInfra (Qwen/Qwen4-Max), read live the same day off
|
|
417
|
+
* api.deepinfra.com/models/Qwen/Qwen4-Max: $1.413/$4.24, cache read
|
|
418
|
+
* 0.0502477× input = $0.071, no write premium, multimodal + input-video
|
|
419
|
+
* tags, and the host's live list carries exactly one qwen4 id (no dated
|
|
420
|
+
* snapshot sibling). Knowledge cutoff: not published by either host —
|
|
421
|
+
* best-effort estimate. qwen4-max supersedes qwen3.8-max: same
|
|
422
|
+
* agentic-flagship tier, cheaper at list on both hosts.)
|
|
377
423
|
* - Zhipu: https://docs.z.ai/guides/overview/pricing + docs.z.ai/guides/llm/
|
|
378
424
|
* glm-5.3 (verified 2026-08-26: glm-5.3 shipped 2026-08-14 and IS in the
|
|
379
425
|
* catalog — $1.40/$4.40, cached $0.26, i.e. glm-5.2's card unchanged, 1M ctx,
|
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;AAwEjD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAmdG;AACH,eAAO,MAAM,MAAM,EAAE,SAAS,eAAe,EAs1FnC,CAAA"}
|
package/dist/models.js
CHANGED
|
@@ -44,8 +44,14 @@ function dateRange(from, to) {
|
|
|
44
44
|
* 2026-09-23). Those windows are 09:00–12:00 and 14:00–18:00 Beijing time, so
|
|
45
45
|
* the Beijing holiday date IS the UTC date of every window.
|
|
46
46
|
*
|
|
47
|
-
*
|
|
48
|
-
* 2026-09-23 at
|
|
47
|
+
* Sources — the State Council's notices, 放假 ranges verbatim:
|
|
48
|
+
* - 2026 (国办发明电〔2025〕7号), read 2026-09-23 at
|
|
49
|
+
* https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html
|
|
50
|
+
* - 2027 (published 2026-11-05), read 2026-11-17 at
|
|
51
|
+
* https://paper.people.com.cn/rmrb/pc/content/202611/06/content_30189674.html
|
|
52
|
+
* (People's Daily reproduction of the 新华社 dispatch; cross-checked against
|
|
53
|
+
* news.ifeng.com/c/8xoAMrtRK7T — all seven clauses and the 2026-11-05 date
|
|
54
|
+
* identical, and every 放假 weekday annotation matches the 2027 calendar).
|
|
49
55
|
* Make-up working days (调休 weekends) need no entry: the peak is Mon–Fri only.
|
|
50
56
|
*
|
|
51
57
|
* ADD NEXT YEAR'S DATES when the State Council publishes them (each November);
|
|
@@ -59,6 +65,13 @@ const CHINA_PUBLIC_HOLIDAYS = [
|
|
|
59
65
|
...dateRange('2026-06-19', '2026-06-21'), // 端午节
|
|
60
66
|
...dateRange('2026-09-25', '2026-09-27'), // 中秋节
|
|
61
67
|
...dateRange('2026-10-01', '2026-10-07'), // 国庆节
|
|
68
|
+
...dateRange('2027-01-01', '2027-01-03'), // 元旦
|
|
69
|
+
...dateRange('2027-02-05', '2027-02-13'), // 春节 (除夕 to 正月初八, 9 days)
|
|
70
|
+
...dateRange('2027-04-03', '2027-04-05'), // 清明节
|
|
71
|
+
...dateRange('2027-05-01', '2027-05-05'), // 劳动节
|
|
72
|
+
...dateRange('2027-06-09', '2027-06-09'), // 端午节 (single day)
|
|
73
|
+
...dateRange('2027-09-15', '2027-09-15'), // 中秋节 (single day)
|
|
74
|
+
...dateRange('2027-10-01', '2027-10-07'), // 国庆节
|
|
62
75
|
];
|
|
63
76
|
/**
|
|
64
77
|
* All available AI models, grouped by provider, ordered from most to least capable.
|
|
@@ -238,6 +251,28 @@ const CHINA_PUBLIC_HOLIDAYS = [
|
|
|
238
251
|
* the Responses API" — the same shape as gpt-6.1-sol, served by the bond's
|
|
239
252
|
* /v1/responses calls. Supersedes gpt-5.6-terra, same tier at $2/$12,
|
|
240
253
|
* repointed from gpt-6.1-sol.)
|
|
254
|
+
* (verified 2026-11-18 on the pricing page + the gpt-6.2-astra model page:
|
|
255
|
+
* ADDED gpt-6.2-astra ($8/$40, cached $0.40 — 5% of input — cache writes
|
|
256
|
+
* $10; >272K band $16/$60, cached $0.80, writes $20 not modeled; 1.05M ctx /
|
|
257
|
+
* 128K out; image input; knowledge cutoff May 18 2026; effort
|
|
258
|
+
* low|medium|high|xhigh|max with NO 'none' — live-probed on /v1/responses
|
|
259
|
+
* with tools at every level 2026-11-18, and the 'none' rejection's error
|
|
260
|
+
* names exactly those five; temperature → 400; forced tool_choice → 200).
|
|
261
|
+
* It supersedes gpt-6-astra (the same flagship tier at $10/$50). The 5.6-sol
|
|
262
|
+
* promo footnote now reads "available until November 30, 2026" — the
|
|
263
|
+
* KNOWN_DIVERGENCES expiry in check-model-freshness.mjs was extended to
|
|
264
|
+
* match. Web search stays $10/1k calls on reasoning models.)
|
|
265
|
+
* (verified 2026-11-18 on the pricing page + the gpt-6.2-sol model page:
|
|
266
|
+
* ADDED gpt-6.2-sol ($2/$10, cached $0.10 — 5% of input — cache writes
|
|
267
|
+
* $2.50; >272K band $4/$15, cached $0.20, writes $5 not modeled; 1.05M ctx /
|
|
268
|
+
* 128K out; image input; knowledge cutoff Jun 1 2026; effort
|
|
269
|
+
* low|medium|high|xhigh|max with NO 'none' — the model page states the API
|
|
270
|
+
* rejects 'none', and all five levels were probed live with tools on
|
|
271
|
+
* /v1/responses 2026-11-18; temperature → 400; forced tool_choice → 200).
|
|
272
|
+
* It supersedes gpt-6.1-sol (same $2/$10 tier, newer generation); gpt-6-sol
|
|
273
|
+
* / 5.6-sol / 5.5 / 5.4 repointed one hop to it. No GPT-6.2 terra or luna id
|
|
274
|
+
* exists on either page. The 5.6-sol promo footnote ("until November 30,
|
|
275
|
+
* 2026") and web-search price are unchanged.)
|
|
241
276
|
* - Google: https://ai.google.dev/gemini-api/docs/pricing (gemini-3.6-flash GA
|
|
242
277
|
* 2026-07-21 $1.50/$7.50 supersedes 3.5-flash as the agentic flagship;
|
|
243
278
|
* gemini-3.1-pro-preview still the pro tier — "3.5 Pro" has NOT shipped as
|
|
@@ -307,11 +342,17 @@ const CHINA_PUBLIC_HOLIDAYS = [
|
|
|
307
342
|
* cataloged superseded by grok-4.7.
|
|
308
343
|
* grok-4.5 still served at $2/$6, cached $0.30; grok-4.3 still served at
|
|
309
344
|
* $1.25/$2.50 with the bigger 1M window; grok-code-fast-1 no longer listed)
|
|
310
|
-
* - Chinese public holidays (DeepSeek's peak excludes them):
|
|
311
|
-
*
|
|
312
|
-
*
|
|
345
|
+
* - Chinese public holidays (DeepSeek's peak excludes them): the State
|
|
346
|
+
* Council's notices — 2026 at
|
|
347
|
+
* https://www.12371.gov.cn/web/article/web/content_1451614684968525824.html,
|
|
348
|
+
* 2027 (published 2026-11-05, read 2026-11-17) at
|
|
349
|
+
* https://paper.people.com.cn/rmrb/pc/content/202611/06/content_30189674.html
|
|
350
|
+
* (see CHINA_PUBLIC_HOLIDAYS)
|
|
313
351
|
* - DeepSeek: https://api-docs.deepseek.com/quick_start/pricing +
|
|
314
|
-
* /updates/ (
|
|
352
|
+
* /updates/ (2026-11-17: the 2027 State Council holiday notice landed above —
|
|
353
|
+
* no DeepSeek card change; this morning's freshness-gate chunk read found the
|
|
354
|
+
* peak-rule sentence and every price unchanged, so the 2026-10-20
|
|
355
|
+
* verification stands. Earlier, verified 2026-10-20 — the whole rate card RE-READ, unchanged:
|
|
315
356
|
* pro 0.66/1.98/0.022 off-peak ×2 peak, flash 0.15/0.6/0.003, peak windows
|
|
316
357
|
* and the Chinese-public-holidays exclusion verbatim as before. The page is
|
|
317
358
|
* now a client-rendered SPA — its HTML is an app shell and the card text
|
|
@@ -426,6 +467,24 @@ const CHINA_PUBLIC_HOLIDAYS = [
|
|
|
426
467
|
* model page still describes it (docs lag the endpoint). The catalog entry
|
|
427
468
|
* dropped its cn region and is now US/DeepInfra-only, whose rates re-verified
|
|
428
469
|
* unchanged the same day: $0.30/$1, cache read 0.3333× input.)
|
|
470
|
+
* (re-verified 2026-11-17 across the provider's own pages: qwen4-max
|
|
471
|
+
* (released 2026-11-16) ADDED. alibabacloud.com/help/en/model-studio/
|
|
472
|
+
* qwen4-max + the EN model-pricing page: Singapore intl $1.5/$4.5 with
|
|
473
|
+
* explicit/implicit cache read $0.075 (5% of input) and every Global region
|
|
474
|
+
* at $1.413/$4.24 with cache read $0.071 — the ZH context-cache doc lists
|
|
475
|
+
* qwen4-max OUTSIDE its standard 20% cache-read table and defers to the
|
|
476
|
+
* console, whose per-region figures the pricing page publishes (explicit
|
|
477
|
+
* 5m creation 125%, 1h 200%; implicit creation 1× input). 1M ctx / 131,072
|
|
478
|
+
* out (262,144 thinking), Text+Image+Video input in every region, hybrid
|
|
479
|
+
* thinking ON by default (thinking_budget, native reasoning_effort
|
|
480
|
+
* low|high|xhigh exists but the bond sends the budget form), parallel tool
|
|
481
|
+
* calling. US = DeepInfra (Qwen/Qwen4-Max), read live the same day off
|
|
482
|
+
* api.deepinfra.com/models/Qwen/Qwen4-Max: $1.413/$4.24, cache read
|
|
483
|
+
* 0.0502477× input = $0.071, no write premium, multimodal + input-video
|
|
484
|
+
* tags, and the host's live list carries exactly one qwen4 id (no dated
|
|
485
|
+
* snapshot sibling). Knowledge cutoff: not published by either host —
|
|
486
|
+
* best-effort estimate. qwen4-max supersedes qwen3.8-max: same
|
|
487
|
+
* agentic-flagship tier, cheaper at list on both hosts.)
|
|
429
488
|
* - Zhipu: https://docs.z.ai/guides/overview/pricing + docs.z.ai/guides/llm/
|
|
430
489
|
* glm-5.3 (verified 2026-08-26: glm-5.3 shipped 2026-08-14 and IS in the
|
|
431
490
|
* catalog — $1.40/$4.40, cached $0.26, i.e. glm-5.2's card unchanged, 1M ctx,
|
|
@@ -1065,7 +1124,58 @@ export const MODELS = [
|
|
|
1065
1124
|
// /v1/responses calls, so it too needs no `toolsRequireReasoningOff` pin.
|
|
1066
1125
|
// At $1/$5 it is 5.6-terra's tier at half the price, so 5.6-terra repoints
|
|
1067
1126
|
// to it.
|
|
1127
|
+
//
|
|
1128
|
+
// gpt-6.2-astra (released 2026-11-17, added 2026-11-18) is astra's own
|
|
1129
|
+
// successor at $8/$40 with cached input at 5% of input (the 6.1-family
|
|
1130
|
+
// ratio, not the 10% GPT-6 carried). Like gpt-6-astra it has no 'none'
|
|
1131
|
+
// effort — the API rejects it and the error names low|medium|high|xhigh|max,
|
|
1132
|
+
// all five verified live with tools on /v1/responses (2026-11-18) —
|
|
1133
|
+
// temperature is rejected like the rest of the family, and a forced
|
|
1134
|
+
// tool_choice returns 200, so it needs no pins at all.
|
|
1135
|
+
//
|
|
1136
|
+
// gpt-6.2-sol (released 2026-11-17, added 2026-11-18) is sol's own successor:
|
|
1137
|
+
// the same $2/$10 list as gpt-6.1-sol with the same 5% cached-input ratio and
|
|
1138
|
+
// $2.50 cache writes, cutoff Jun 1 2026. Identical probe shape to
|
|
1139
|
+
// gpt-6.2-astra: no 'none' effort (the rejection's error names exactly
|
|
1140
|
+
// low|medium|high|xhigh|max), all five levels verified live with tools on
|
|
1141
|
+
// /v1/responses (2026-11-18), temperature → 400, forced tool_choice → 200 —
|
|
1142
|
+
// no pins. It supersedes gpt-6.1-sol (same tier, same price, newer
|
|
1143
|
+
// generation); no GPT-6.2 terra or luna id exists.
|
|
1068
1144
|
// ---------------------------------------------------------------------------
|
|
1145
|
+
{
|
|
1146
|
+
id: 'gpt-6.2-astra',
|
|
1147
|
+
provider: 'openai',
|
|
1148
|
+
label: 'GPT-6.2 Astra',
|
|
1149
|
+
description: 'OpenAI flagship — the hardest reasoning & coding work',
|
|
1150
|
+
// Documented as 1.05M; floored to 1M like the other OpenAI entries.
|
|
1151
|
+
contextWindow: 1_000_000,
|
|
1152
|
+
maxOutputTokens: 128_000,
|
|
1153
|
+
supportsThinking: true,
|
|
1154
|
+
thinkingBudgetTokens: 16_000,
|
|
1155
|
+
thinkingConfigurable: true,
|
|
1156
|
+
// No 'none' on this model. All five levels verified accepted with tools on
|
|
1157
|
+
// /v1/responses (2026-11-18) — the 'none' rejection's own error message
|
|
1158
|
+
// lists exactly these values.
|
|
1159
|
+
supportedEffortLevels: ['low', 'medium', 'high', 'xhigh', 'max'],
|
|
1160
|
+
defaultEffortLevel: 'medium',
|
|
1161
|
+
// temperature → 400 "Unsupported parameter: 'temperature' is not supported
|
|
1162
|
+
// with this model" (probed live 2026-11-18).
|
|
1163
|
+
rejectsTemperature: true,
|
|
1164
|
+
supportsVision: true,
|
|
1165
|
+
supportsPromptCaching: true,
|
|
1166
|
+
supportsTools: true,
|
|
1167
|
+
webSearchToolType: 'web_search',
|
|
1168
|
+
// $10 per 1,000 web_search_call items (pricing page tool table, reasoning
|
|
1169
|
+
// models); search-result tokens bill as input.
|
|
1170
|
+
webSearchPricePer1k: 10,
|
|
1171
|
+
codeExecutionToolType: 'code_interpreter',
|
|
1172
|
+
// Standard tier; >272K band ($16/$60, cached $0.80, writes $20) not modeled.
|
|
1173
|
+
inputPricePerMTok: 8,
|
|
1174
|
+
outputPricePerMTok: 40,
|
|
1175
|
+
cacheReadPricePerMTok: 0.4,
|
|
1176
|
+
cacheWritePricePerMTok: 10,
|
|
1177
|
+
knowledgeCutoff: '2026-05-18',
|
|
1178
|
+
},
|
|
1069
1179
|
{
|
|
1070
1180
|
id: 'gpt-6-astra',
|
|
1071
1181
|
provider: 'openai',
|
|
@@ -1098,6 +1208,47 @@ export const MODELS = [
|
|
|
1098
1208
|
cacheReadPricePerMTok: 1,
|
|
1099
1209
|
cacheWritePricePerMTok: 12.5,
|
|
1100
1210
|
knowledgeCutoff: '2026-04-30',
|
|
1211
|
+
// Superseded by gpt-6.2-astra — the same flagship tier at $8/$40 with a
|
|
1212
|
+
// fifth of the cached-input rate. Still served and priceable — just not
|
|
1213
|
+
// offered in the picker.
|
|
1214
|
+
deprecatedAt: '2026-11-18',
|
|
1215
|
+
supersededBy: 'gpt-6.2-astra',
|
|
1216
|
+
},
|
|
1217
|
+
{
|
|
1218
|
+
id: 'gpt-6.2-sol',
|
|
1219
|
+
provider: 'openai',
|
|
1220
|
+
label: 'GPT-6.2 Sol',
|
|
1221
|
+
description: 'OpenAI for complex coding & agentic work',
|
|
1222
|
+
// Documented as 1.05M; floored to 1M like the other OpenAI entries.
|
|
1223
|
+
contextWindow: 1_000_000,
|
|
1224
|
+
maxOutputTokens: 128_000,
|
|
1225
|
+
supportsThinking: true,
|
|
1226
|
+
thinkingBudgetTokens: 16_000,
|
|
1227
|
+
thinkingConfigurable: true,
|
|
1228
|
+
// No 'none' on this model (model page: "The API rejects requests that set
|
|
1229
|
+
// reasoning effort to `none`."). All five levels verified accepted with
|
|
1230
|
+
// tools on /v1/responses (2026-11-18) — the 'none' rejection's own error
|
|
1231
|
+
// message lists exactly these values.
|
|
1232
|
+
supportedEffortLevels: ['low', 'medium', 'high', 'xhigh', 'max'],
|
|
1233
|
+
defaultEffortLevel: 'medium',
|
|
1234
|
+
// temperature → 400 "Unsupported parameter: 'temperature' is not supported
|
|
1235
|
+
// with this model" (probed live 2026-11-18).
|
|
1236
|
+
rejectsTemperature: true,
|
|
1237
|
+
supportsVision: true,
|
|
1238
|
+
supportsPromptCaching: true,
|
|
1239
|
+
supportsTools: true,
|
|
1240
|
+
webSearchToolType: 'web_search',
|
|
1241
|
+
// $10 per 1,000 web_search_call items (pricing page tool table, reasoning
|
|
1242
|
+
// models); search-result tokens bill as input.
|
|
1243
|
+
webSearchPricePer1k: 10,
|
|
1244
|
+
codeExecutionToolType: 'code_interpreter',
|
|
1245
|
+
// Standard tier. Cached input is 5% of input like the 6.1 family; >272K
|
|
1246
|
+
// band ($4/$15, cached $0.20, writes $5) not modeled.
|
|
1247
|
+
inputPricePerMTok: 2,
|
|
1248
|
+
outputPricePerMTok: 10,
|
|
1249
|
+
cacheReadPricePerMTok: 0.1,
|
|
1250
|
+
cacheWritePricePerMTok: 2.5,
|
|
1251
|
+
knowledgeCutoff: '2026-06-01',
|
|
1101
1252
|
},
|
|
1102
1253
|
{
|
|
1103
1254
|
id: 'gpt-6.1-sol',
|
|
@@ -1129,6 +1280,11 @@ export const MODELS = [
|
|
|
1129
1280
|
cacheReadPricePerMTok: 0.1,
|
|
1130
1281
|
cacheWritePricePerMTok: 2.5,
|
|
1131
1282
|
knowledgeCutoff: '2026-04-30',
|
|
1283
|
+
// Superseded by gpt-6.2-sol — the same $2/$10 tier with a newer
|
|
1284
|
+
// generation (cutoff Jun 1 2026 vs Apr 30 2026). Still served and
|
|
1285
|
+
// priceable — just not offered in the picker.
|
|
1286
|
+
deprecatedAt: '2026-11-18',
|
|
1287
|
+
supersededBy: 'gpt-6.2-sol',
|
|
1132
1288
|
},
|
|
1133
1289
|
{
|
|
1134
1290
|
id: 'gpt-6.1-terra',
|
|
@@ -1195,10 +1351,11 @@ export const MODELS = [
|
|
|
1195
1351
|
cacheReadPricePerMTok: 0.2,
|
|
1196
1352
|
cacheWritePricePerMTok: 2.5,
|
|
1197
1353
|
knowledgeCutoff: '2026-04-20',
|
|
1198
|
-
// Superseded by gpt-6.1-sol
|
|
1199
|
-
// Still served and priceable — just not offered in the
|
|
1354
|
+
// Superseded by gpt-6.2-sol (via gpt-6.1-sol: same $2/$10, cached input
|
|
1355
|
+
// $0.10 vs $0.20). Still served and priceable — just not offered in the
|
|
1356
|
+
// picker.
|
|
1200
1357
|
deprecatedAt: '2026-09-30',
|
|
1201
|
-
supersededBy: 'gpt-6.
|
|
1358
|
+
supersededBy: 'gpt-6.2-sol',
|
|
1202
1359
|
},
|
|
1203
1360
|
{
|
|
1204
1361
|
id: 'gpt-6-luna',
|
|
@@ -1260,10 +1417,11 @@ export const MODELS = [
|
|
|
1260
1417
|
codeExecutionToolType: 'code_interpreter',
|
|
1261
1418
|
// LIST price. OpenAI ran a >20% PROMO from 2026-08-22 ($4/$20, cache read
|
|
1262
1419
|
// $0.40, cache write $5) — "GPT-5.6 Sol's promotional pricing is available
|
|
1263
|
-
//
|
|
1420
|
+
// until November 30, 2026" per the pricing page's own footnote (re-read
|
|
1421
|
+
// 2026-11-18; it previously said "at least through November 21, 2026").
|
|
1264
1422
|
// Billed here at list so metering never under-charges when it lapses; same
|
|
1265
1423
|
// policy as the gemini-flash and claude-sonnet-5 promos, and recorded in
|
|
1266
|
-
// check-model-freshness.mjs KNOWN_DIVERGENCES (expires 2026-11-
|
|
1424
|
+
// check-model-freshness.mjs KNOWN_DIVERGENCES (expires 2026-11-30).
|
|
1267
1425
|
inputPricePerMTok: 5,
|
|
1268
1426
|
outputPricePerMTok: 30,
|
|
1269
1427
|
// Cached input 0.1× input, cache write 1.25× input (see section note).
|
|
@@ -1272,10 +1430,10 @@ export const MODELS = [
|
|
|
1272
1430
|
// Not published — best-effort estimate.
|
|
1273
1431
|
knowledgeCutoff: '2026-03-01',
|
|
1274
1432
|
// Superseded by gpt-6-sol ($2/$10 vs $5/$30 list); points one hop to
|
|
1275
|
-
// gpt-6.
|
|
1276
|
-
// priceable — just not offered in the picker.
|
|
1433
|
+
// gpt-6.2-sol since 2026-11-18, when gpt-6.1-sol was itself superseded.
|
|
1434
|
+
// Still served and priceable — just not offered in the picker.
|
|
1277
1435
|
deprecatedAt: '2026-09-23',
|
|
1278
|
-
supersededBy: 'gpt-6.
|
|
1436
|
+
supersededBy: 'gpt-6.2-sol',
|
|
1279
1437
|
},
|
|
1280
1438
|
{
|
|
1281
1439
|
id: 'gpt-5.6-terra',
|
|
@@ -1389,11 +1547,11 @@ export const MODELS = [
|
|
|
1389
1547
|
cacheWritePricePerMTok: 5,
|
|
1390
1548
|
knowledgeCutoff: '2025-12-01',
|
|
1391
1549
|
// Superseded by gpt-5.6-sol (same frontier tier, same $5/$30), and since
|
|
1392
|
-
// 2026-09-23 points one hop to
|
|
1393
|
-
// since 2026-
|
|
1394
|
-
//
|
|
1550
|
+
// 2026-09-23 points one hop to the sol line's current generation (gpt-6.2-sol
|
|
1551
|
+
// since 2026-11-18). Still listed as current by OpenAI, so it stays
|
|
1552
|
+
// priceable — just not offered.
|
|
1395
1553
|
deprecatedAt: '2026-07-09',
|
|
1396
|
-
supersededBy: 'gpt-6.
|
|
1554
|
+
supersededBy: 'gpt-6.2-sol',
|
|
1397
1555
|
},
|
|
1398
1556
|
{
|
|
1399
1557
|
id: 'gpt-5.4',
|
|
@@ -1427,9 +1585,9 @@ export const MODELS = [
|
|
|
1427
1585
|
// offers only the 5.6 generation (this is OUR taxonomy, not OpenAI's
|
|
1428
1586
|
// deprecations page; the model stays priceable). Points one hop to
|
|
1429
1587
|
// gpt-6-sol since 2026-09-23, when 5.6-terra was itself superseded, and to
|
|
1430
|
-
//
|
|
1588
|
+
// the sol line's current generation since 2026-11-18.
|
|
1431
1589
|
deprecatedAt: '2026-07-28',
|
|
1432
|
-
supersededBy: 'gpt-6.
|
|
1590
|
+
supersededBy: 'gpt-6.2-sol',
|
|
1433
1591
|
},
|
|
1434
1592
|
{
|
|
1435
1593
|
id: 'gpt-5.4-mini',
|
|
@@ -2796,20 +2954,78 @@ export const MODELS = [
|
|
|
2796
2954
|
// Verified: https://www.alibabacloud.com/help/en/model-studio/deep-thinking
|
|
2797
2955
|
// https://www.alibabacloud.com/help/en/model-studio/qwen-coder
|
|
2798
2956
|
// https://openrouter.ai/qwen/qwen3.7-max
|
|
2799
|
-
//
|
|
2800
|
-
//
|
|
2801
|
-
//
|
|
2802
|
-
//
|
|
2803
|
-
//
|
|
2804
|
-
//
|
|
2957
|
+
// qwen4-max (2026-11-16) is the agentic flagship, succeeding qwen3.8-max —
|
|
2958
|
+
// multimodal (Text+Image+Video input) on BOTH hosts this generation. Hybrid
|
|
2959
|
+
// thinking uses enable_thinking (default ON) + thinking_budget (token cap) —
|
|
2960
|
+
// a real budget param, so effort scales the budget. (Qwen4-Max also exposes a
|
|
2961
|
+
// native reasoning_effort low|high|xhigh, but the alibaba bond speaks only
|
|
2962
|
+
// the budget form, so the catalog models effort as scaled budgets like the
|
|
2963
|
+
// 3.x flagships.) The qwen3-coder models are NON-thinking.
|
|
2805
2964
|
// Prices are DashScope international list rates (the bond calls DashScope,
|
|
2806
|
-
// not OpenRouter;
|
|
2965
|
+
// not OpenRouter; bill at list).
|
|
2807
2966
|
// ---------------------------------------------------------------------------
|
|
2808
|
-
//
|
|
2967
|
+
// qwen4-max (2026-11-16) succeeds qwen3.8-max as the agentic flagship,
|
|
2968
|
+
// priced BELOW it at $1.5/$4.5 on the Singapore international card (every
|
|
2969
|
+
// other native region and the US re-host bill $1.413/$4.24). Same hybrid
|
|
2970
|
+
// thinking mechanism as the 3.x series. Cache rates are published per region
|
|
2971
|
+
// on the model-pricing page — see the Sources block for the citation.
|
|
2972
|
+
{
|
|
2973
|
+
id: 'qwen4-max',
|
|
2974
|
+
provider: 'alibaba',
|
|
2975
|
+
label: 'Qwen4 Max',
|
|
2976
|
+
description: 'Alibaba agentic flagship — 1M context, multimodal, hybrid thinking',
|
|
2977
|
+
// Model page (verified 2026-11-17): context 1,000,000; max input 991,808
|
|
2978
|
+
// (983,616 thinking); max output 131,072 (262,144 in thinking mode); max
|
|
2979
|
+
// chain-of-thought 262,144. Kept at the non-thinking output figure —
|
|
2980
|
+
// understating only shortens completions.
|
|
2981
|
+
contextWindow: 1_000_000,
|
|
2982
|
+
maxOutputTokens: 131_072,
|
|
2983
|
+
supportsThinking: true,
|
|
2984
|
+
thinkingBudgetTokens: 8_000,
|
|
2985
|
+
thinkingConfigurable: true,
|
|
2986
|
+
supportedEffortLevels: ['4K', '8K', '16K', '32K'],
|
|
2987
|
+
defaultEffortLevel: '8K',
|
|
2988
|
+
effortBudgetTokens: { '4K': 4000, '8K': 8000, '16K': 16000, '32K': 32000 },
|
|
2989
|
+
// Multimodal on BOTH hosts this generation: the model page lists
|
|
2990
|
+
// Text/Image/Video input in every region, and the DeepInfra re-host's own
|
|
2991
|
+
// record tags the id multimodal + input-video (Qwen3.8-Max's re-host
|
|
2992
|
+
// carried no such tags — that is why vision stayed false there).
|
|
2993
|
+
supportsVision: true,
|
|
2994
|
+
supportsPromptCaching: true,
|
|
2995
|
+
supportsTools: true,
|
|
2996
|
+
// Singapore international list card (model page + model-pricing page,
|
|
2997
|
+
// verified 2026-11-17). The ZH context-cache doc lists qwen4-max among the
|
|
2998
|
+
// models OUTSIDE the standard 20% cache-read table and defers to the
|
|
2999
|
+
// console, whose figures the pricing page publishes per region: cache read
|
|
3000
|
+
// $0.075 in Singapore (5% of input; $0.071 in every Global region).
|
|
3001
|
+
// Implicit creation is billed at plain input (write 1×); the explicit 5m
|
|
3002
|
+
// creation premium is 125% = $1.875, which implicit-mode billing never
|
|
3003
|
+
// charges.
|
|
3004
|
+
inputPricePerMTok: 1.5,
|
|
3005
|
+
outputPricePerMTok: 4.5,
|
|
3006
|
+
cacheReadPricePerMTok: 0.075,
|
|
3007
|
+
cacheWritePricePerMTok: 1.5,
|
|
3008
|
+
regions: ['us', 'cn'],
|
|
3009
|
+
// US = DeepInfra (Qwen/Qwen4-Max), verified 2026-11-17 against
|
|
3010
|
+
// api.deepinfra.com/models/Qwen/Qwen4-Max: $1.413/$4.24, cache read
|
|
3011
|
+
// 0.0502477× input = $0.071, no write premium (omitted → region input).
|
|
3012
|
+
// The host's live model list carries exactly one qwen4 id — no dated
|
|
3013
|
+
// snapshot sibling — and the entry is not deprecated.
|
|
3014
|
+
regionPricing: {
|
|
3015
|
+
us: { inputPricePerMTok: 1.413, outputPricePerMTok: 4.24, cacheReadPricePerMTok: 0.071 },
|
|
3016
|
+
},
|
|
3017
|
+
// Not published by Alibaba — best-effort estimate for the Qwen4 generation
|
|
3018
|
+
// (released 2026-11-16).
|
|
3019
|
+
knowledgeCutoff: '2026-08-01',
|
|
3020
|
+
},
|
|
3021
|
+
// qwen3.8-max (GA 2026-08-03) succeeded qwen3.7-max as the agentic flagship,
|
|
2809
3022
|
// priced BELOW it at $2/$6 (intl CNY 14.988/44.965, same fixed conversion).
|
|
2810
3023
|
// Same hybrid thinking mechanism as the 3.7 series (enable_thinking default
|
|
2811
3024
|
// ON + thinking_budget; preserve_thinking supported). Context cache uses the
|
|
2812
3025
|
// standard implicit rates — see the Sources block for the ZH-doc citation.
|
|
3026
|
+
// Superseded by qwen4-max (2026-11-16): same tier at a LOWER list price on
|
|
3027
|
+
// both hosts. Still served upstream, so it stays priceable — just not
|
|
3028
|
+
// offered.
|
|
2813
3029
|
{
|
|
2814
3030
|
id: 'qwen3.8-max',
|
|
2815
3031
|
provider: 'alibaba',
|
|
@@ -2856,6 +3072,8 @@ export const MODELS = [
|
|
|
2856
3072
|
},
|
|
2857
3073
|
// Not published by Alibaba — best-effort estimate.
|
|
2858
3074
|
knowledgeCutoff: '2026-04-01',
|
|
3075
|
+
deprecatedAt: '2026-11-16',
|
|
3076
|
+
supersededBy: 'qwen4-max',
|
|
2859
3077
|
},
|
|
2860
3078
|
{
|
|
2861
3079
|
id: 'qwen3.8-flash',
|
|
@@ -2933,11 +3151,13 @@ export const MODELS = [
|
|
|
2933
3151
|
},
|
|
2934
3152
|
// Not published by Alibaba — best-effort estimate.
|
|
2935
3153
|
knowledgeCutoff: '2026-01-01',
|
|
2936
|
-
// Superseded
|
|
2937
|
-
//
|
|
2938
|
-
//
|
|
3154
|
+
// Superseded (GA 2026-08-03) and now pointed one hop forward at qwen4-max
|
|
3155
|
+
// (2026-11-16): same tier and mechanism, each generation CHEAPER at list
|
|
3156
|
+
// ($2/$6 then $1.5/$4.5, vs this id's $2.50/$7.50). Still served upstream
|
|
3157
|
+
// (the 50%-off promo runs on this id), so it stays priceable — it is just
|
|
3158
|
+
// not offered.
|
|
2939
3159
|
deprecatedAt: '2026-08-03',
|
|
2940
|
-
supersededBy: '
|
|
3160
|
+
supersededBy: 'qwen4-max',
|
|
2941
3161
|
},
|
|
2942
3162
|
{
|
|
2943
3163
|
id: 'qwen3-coder-plus',
|
package/package.json
CHANGED
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@molecule/api-resource-ai-models",
|
|
3
|
-
"version": "1.
|
|
4
|
-
"description": "AI model catalog
|
|
3
|
+
"version": "1.22.0",
|
|
4
|
+
"description": "Server-side AI model catalog plus GET /ai/models endpoint",
|
|
5
5
|
"type": "module",
|
|
6
6
|
"main": "dist/index.js",
|
|
7
7
|
"types": "dist/index.d.ts",
|
|
@@ -28,8 +28,8 @@
|
|
|
28
28
|
"license": "Apache-2.0",
|
|
29
29
|
"devDependencies": {
|
|
30
30
|
"@molecule/api-bond": "1.0.4",
|
|
31
|
-
"@molecule/api-i18n": "1.0.
|
|
32
|
-
"@molecule/api-resource": "1.1.
|
|
31
|
+
"@molecule/api-i18n": "1.0.6",
|
|
32
|
+
"@molecule/api-resource": "1.1.3",
|
|
33
33
|
"@types/node": "26.6.5",
|
|
34
34
|
"typescript": "6.0.3",
|
|
35
35
|
"vitest": "4.1.12"
|