@nodaro/shared 3.62.0 → 3.64.0
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/dist/{chunk-C7POAYFE.js → chunk-36VWLB4J.js} +3 -3
- package/dist/{chunk-C7POAYFE.js.map → chunk-36VWLB4J.js.map} +1 -1
- package/dist/{chunk-S24ZBUB4.js → chunk-3AYM47GW.js} +6 -32
- package/dist/chunk-3AYM47GW.js.map +1 -0
- package/dist/{chunk-YCXRSWCQ.js → chunk-3KTWYYCN.js} +8 -5
- package/dist/chunk-3KTWYYCN.js.map +1 -0
- package/dist/{chunk-ZES6EZC4.js → chunk-4IUE2JB7.js} +3 -3
- package/dist/{chunk-ZES6EZC4.js.map → chunk-4IUE2JB7.js.map} +1 -1
- package/dist/{chunk-MYJQELFN.cjs → chunk-5OBTGNUQ.cjs} +547 -29
- package/dist/chunk-5OBTGNUQ.cjs.map +1 -0
- package/dist/{chunk-TWF2OSH4.cjs → chunk-67CMMTLN.cjs} +4 -4
- package/dist/{chunk-TWF2OSH4.cjs.map → chunk-67CMMTLN.cjs.map} +1 -1
- package/dist/{chunk-WRSWH7US.cjs → chunk-6BRWPJPP.cjs} +17 -2
- package/dist/chunk-6BRWPJPP.cjs.map +1 -0
- package/dist/{chunk-S5TMRJMX.js → chunk-6YIV37V7.js} +4 -4
- package/dist/{chunk-S5TMRJMX.js.map → chunk-6YIV37V7.js.map} +1 -1
- package/dist/{chunk-3M4BNKFJ.cjs → chunk-7AJGUI7L.cjs} +202 -16
- package/dist/chunk-7AJGUI7L.cjs.map +1 -0
- package/dist/{chunk-JZGZ26YB.js → chunk-7KQ6W5DS.js} +4 -4
- package/dist/{chunk-JZGZ26YB.js.map → chunk-7KQ6W5DS.js.map} +1 -1
- package/dist/{chunk-IWWVPR3W.cjs → chunk-ACZZHHFZ.cjs} +149 -34
- package/dist/chunk-ACZZHHFZ.cjs.map +1 -0
- package/dist/{chunk-2W745U4A.cjs → chunk-AZNZ5WNZ.cjs} +42 -70
- package/dist/chunk-AZNZ5WNZ.cjs.map +1 -0
- package/dist/{chunk-HSHCCBIV.js → chunk-BQYCZLMF.js} +16 -3
- package/dist/chunk-BQYCZLMF.js.map +1 -0
- package/dist/chunk-CHS2SIP6.cjs +504 -0
- package/dist/chunk-CHS2SIP6.cjs.map +1 -0
- package/dist/{chunk-CIPPPT7E.cjs → chunk-CHTZZRAL.cjs} +6 -6
- package/dist/{chunk-CIPPPT7E.cjs.map → chunk-CHTZZRAL.cjs.map} +1 -1
- package/dist/{chunk-AOJ6B7K4.cjs → chunk-D4KWNDPH.cjs} +12 -12
- package/dist/chunk-D4KWNDPH.cjs.map +1 -0
- package/dist/{chunk-5YYN5X57.js → chunk-DZ4IOP3Q.js} +536 -30
- package/dist/chunk-DZ4IOP3Q.js.map +1 -0
- package/dist/{chunk-VACPZPUA.cjs → chunk-EXEVT6D2.cjs} +15 -5
- package/dist/chunk-EXEVT6D2.cjs.map +1 -0
- package/dist/{chunk-5M43XMRB.cjs → chunk-FHGRKQVZ.cjs} +8 -8
- package/dist/{chunk-5M43XMRB.cjs.map → chunk-FHGRKQVZ.cjs.map} +1 -1
- package/dist/{chunk-KJSRPHJ2.cjs → chunk-GTWUS6MQ.cjs} +165 -54
- package/dist/chunk-GTWUS6MQ.cjs.map +1 -0
- package/dist/{chunk-Q6IES3U2.js → chunk-H5KHWTAE.js} +5 -2
- package/dist/chunk-H5KHWTAE.js.map +1 -0
- package/dist/{chunk-G3KVK7OU.cjs → chunk-HDPJNNMF.cjs} +21 -21
- package/dist/{chunk-G3KVK7OU.cjs.map → chunk-HDPJNNMF.cjs.map} +1 -1
- package/dist/{chunk-3OYO3XCC.js → chunk-IYSOVS7O.js} +118 -24
- package/dist/chunk-IYSOVS7O.js.map +1 -0
- package/dist/{chunk-AFYTX4AI.cjs → chunk-J6B7J73J.cjs} +110 -50
- package/dist/chunk-J6B7J73J.cjs.map +1 -0
- package/dist/{chunk-I2RQB7VG.js → chunk-KB2SQAR3.js} +3 -3
- package/dist/{chunk-I2RQB7VG.js.map → chunk-KB2SQAR3.js.map} +1 -1
- package/dist/{chunk-BLJUCPMV.cjs → chunk-KSHAAPQF.cjs} +5 -2
- package/dist/chunk-KSHAAPQF.cjs.map +1 -0
- package/dist/{chunk-6G7NN4S4.js → chunk-NIAFR5EE.js} +2 -2
- package/dist/chunk-NIAFR5EE.js.map +1 -0
- package/dist/{chunk-CF4S4N5X.js → chunk-O2QOBA6Q.js} +2 -2
- package/dist/{chunk-CF4S4N5X.js.map → chunk-O2QOBA6Q.js.map} +1 -1
- package/dist/{chunk-IDTKF5EX.cjs → chunk-O4RPXBDZ.cjs} +12 -8
- package/dist/chunk-O4RPXBDZ.cjs.map +1 -0
- package/dist/{chunk-4GGLKRYI.cjs → chunk-O6TBB4IF.cjs} +2 -2
- package/dist/{chunk-4GGLKRYI.cjs.map → chunk-O6TBB4IF.cjs.map} +1 -1
- package/dist/{chunk-PTD6EDUU.js → chunk-OBOOHG3N.js} +13 -4
- package/dist/chunk-OBOOHG3N.js.map +1 -0
- package/dist/{chunk-2D5LQFV4.cjs → chunk-OOZW2MM6.cjs} +9 -9
- package/dist/{chunk-2D5LQFV4.cjs.map → chunk-OOZW2MM6.cjs.map} +1 -1
- package/dist/{chunk-A2PX5IRC.cjs → chunk-PERUFUVH.cjs} +13 -13
- package/dist/{chunk-A2PX5IRC.cjs.map → chunk-PERUFUVH.cjs.map} +1 -1
- package/dist/{chunk-G4U3HW6W.js → chunk-PJJAZQVZ.js} +3 -3
- package/dist/{chunk-G4U3HW6W.js.map → chunk-PJJAZQVZ.js.map} +1 -1
- package/dist/{chunk-E2HZTGYX.js → chunk-Q2Y4JHF7.js} +7 -7
- package/dist/chunk-Q2Y4JHF7.js.map +1 -0
- package/dist/{chunk-FVII3KHN.js → chunk-QYU7J5EQ.js} +198 -17
- package/dist/chunk-QYU7J5EQ.js.map +1 -0
- package/dist/{chunk-JJNNM23P.js → chunk-S3UGLZSF.js} +3 -3
- package/dist/{chunk-JJNNM23P.js.map → chunk-S3UGLZSF.js.map} +1 -1
- package/dist/{chunk-CC4V2B7S.js → chunk-SHH2T4CR.js} +3 -3
- package/dist/{chunk-CC4V2B7S.js.map → chunk-SHH2T4CR.js.map} +1 -1
- package/dist/{chunk-4ZEFP2FQ.cjs → chunk-SPAHI7DY.cjs} +4 -4
- package/dist/{chunk-4ZEFP2FQ.cjs.map → chunk-SPAHI7DY.cjs.map} +1 -1
- package/dist/{chunk-Y6X7UJSD.js → chunk-SXHHLZ22.js} +122 -15
- package/dist/chunk-SXHHLZ22.js.map +1 -0
- package/dist/{chunk-HRY7KGAE.cjs → chunk-T2OVDF2M.cjs} +6 -6
- package/dist/{chunk-HRY7KGAE.cjs.map → chunk-T2OVDF2M.cjs.map} +1 -1
- package/dist/{chunk-2QKF7M2Z.js → chunk-TAPBDT47.js} +3 -3
- package/dist/{chunk-2QKF7M2Z.js.map → chunk-TAPBDT47.js.map} +1 -1
- package/dist/{chunk-NGKKNSK6.js → chunk-TPQJKS7D.js} +139 -16
- package/dist/chunk-TPQJKS7D.js.map +1 -0
- package/dist/{chunk-N7OUQK7E.cjs → chunk-TQW3R44P.cjs} +8 -8
- package/dist/{chunk-N7OUQK7E.cjs.map → chunk-TQW3R44P.cjs.map} +1 -1
- package/dist/{chunk-GSFB537T.js → chunk-TTGZ2L3O.js} +23 -4
- package/dist/chunk-TTGZ2L3O.js.map +1 -0
- package/dist/{chunk-SYWR5G3T.js → chunk-VPTA4Z7G.js} +54 -7
- package/dist/chunk-VPTA4Z7G.js.map +1 -0
- package/dist/{chunk-2HZGDZZN.cjs → chunk-VZ6TE23U.cjs} +6 -6
- package/dist/{chunk-2HZGDZZN.cjs.map → chunk-VZ6TE23U.cjs.map} +1 -1
- package/dist/{chunk-TRVWMKNE.cjs → chunk-W6VJAUGM.cjs} +285 -64
- package/dist/chunk-W6VJAUGM.cjs.map +1 -0
- package/dist/{chunk-UDDVDQMF.js → chunk-W7KHHNZR.js} +235 -37
- package/dist/chunk-W7KHHNZR.js.map +1 -0
- package/dist/{chunk-KYMHFVAZ.cjs → chunk-Y4GCJJPF.cjs} +2 -2
- package/dist/chunk-Y4GCJJPF.cjs.map +1 -0
- package/dist/{chunk-6CSXMWL4.cjs → chunk-YQV53HRR.cjs} +27 -4
- package/dist/chunk-YQV53HRR.cjs.map +1 -0
- package/dist/{chunk-7FCTCWHD.js → chunk-ZI3QDUHO.js} +3 -3
- package/dist/{chunk-7FCTCWHD.js.map → chunk-ZI3QDUHO.js.map} +1 -1
- package/dist/entries/apps.cjs +43 -42
- package/dist/entries/apps.js +8 -7
- package/dist/entries/captions.cjs +38 -30
- package/dist/entries/captions.d.cts +47 -1
- package/dist/entries/captions.d.ts +47 -1
- package/dist/entries/captions.js +1 -1
- package/dist/entries/credits.cjs +354 -174
- package/dist/entries/credits.d.cts +378 -27
- package/dist/entries/credits.d.ts +378 -27
- package/dist/entries/credits.js +11 -11
- package/dist/entries/llm-models.cjs +77 -57
- package/dist/entries/llm-models.d.cts +108 -7
- package/dist/entries/llm-models.d.ts +108 -7
- package/dist/entries/llm-models.js +2 -2
- package/dist/entries/media.cjs +63 -63
- package/dist/entries/media.d.cts +8 -0
- package/dist/entries/media.d.ts +8 -0
- package/dist/entries/media.js +1 -1
- package/dist/entries/model-catalog.cjs +161 -105
- package/dist/entries/model-catalog.d.cts +109 -22
- package/dist/entries/model-catalog.d.ts +109 -22
- package/dist/entries/model-catalog.js +4 -4
- package/dist/entries/models.cjs +466 -322
- package/dist/entries/models.d.cts +71 -5
- package/dist/entries/models.d.ts +71 -5
- package/dist/entries/models.js +5 -5
- package/dist/entries/nodes.cjs +226 -209
- package/dist/entries/nodes.d.cts +66 -7
- package/dist/entries/nodes.d.ts +66 -7
- package/dist/entries/nodes.js +10 -9
- package/dist/entries/pipeline.cjs +156 -152
- package/dist/entries/pipeline.d.cts +14 -4
- package/dist/entries/pipeline.d.ts +14 -4
- package/dist/entries/pipeline.js +3 -3
- package/dist/entries/runtime-config.cjs +29 -29
- package/dist/entries/runtime-config.d.cts +10 -7
- package/dist/entries/runtime-config.d.ts +10 -7
- package/dist/entries/runtime-config.js +1 -1
- package/dist/entries/social.cjs +177 -177
- package/dist/entries/social.js +2 -2
- package/dist/entries/types.d.cts +8 -0
- package/dist/entries/types.d.ts +8 -0
- package/dist/entries/ugc.cjs +71 -71
- package/dist/entries/ugc.js +4 -4
- package/dist/entries/video.cjs +154 -150
- package/dist/entries/video.d.cts +11 -2
- package/dist/entries/video.d.ts +11 -2
- package/dist/entries/video.js +6 -6
- package/dist/entries/workflow.cjs +150 -150
- package/dist/entries/workflow.d.cts +1 -1
- package/dist/entries/workflow.d.ts +1 -1
- package/dist/entries/workflow.js +2 -2
- package/dist/index.cjs +2136 -1704
- package/dist/index.d.cts +10 -10
- package/dist/index.d.ts +10 -10
- package/dist/index.js +27 -27
- package/dist/{model-catalog-CCEcv08q.d.ts → model-catalog-Cf9fkEUp.d.cts} +45 -2
- package/dist/{model-catalog-CCEcv08q.d.cts → model-catalog-Cf9fkEUp.d.ts} +45 -2
- package/dist/{model-constants-B88p6XR0.d.ts → model-constants-BxjiWVQ9.d.ts} +215 -32
- package/dist/{model-constants-Cf3zaITt.d.cts → model-constants-C_Nyft_i.d.cts} +215 -32
- package/package.json +1 -1
- package/src/__tests__/add-captions-nemotron-pricing.test.ts +111 -0
- package/src/__tests__/credit-identifiers.test.ts +29 -0
- package/src/__tests__/extend-providers.test.ts +359 -39
- package/src/__tests__/fixtures/mask-composite-quote-parity.json +1 -1
- package/src/__tests__/fixtures/measured-input-quote-parity.json +1435 -0
- package/src/__tests__/fixtures/planned-length-quote-parity.json +1866 -0
- package/src/__tests__/fixtures/video-utility-quote-parity.json +2343 -0
- package/src/__tests__/flux-3-klein.test.ts +203 -0
- package/src/__tests__/gpt-image-2-6-formats.test.ts +3 -1
- package/src/__tests__/gvp-supported-providers.test.ts +27 -2
- package/src/__tests__/lip-sync-audio-rate.test.ts +74 -0
- package/src/__tests__/lip-sync-levers.test.ts +31 -0
- package/src/__tests__/llm-models.test.ts +139 -12
- package/src/__tests__/llm-price-key.test.ts +164 -1
- package/src/__tests__/model-makers.test.ts +3 -0
- package/src/__tests__/nemotron-transcribe-pricing.test.ts +61 -0
- package/src/__tests__/node-runtime-keys.test.ts +7 -0
- package/src/__tests__/normalize-node-params.test.ts +172 -1
- package/src/__tests__/pixverse-v6-max-catalog.test.ts +271 -0
- package/src/__tests__/prompt-length-limits.test.ts +4 -0
- package/src/__tests__/retired-lip-sync-providers.test.ts +43 -0
- package/src/__tests__/seedance-extend-source-credit-id.test.ts +89 -0
- package/src/__tests__/seedance-task-type.test.ts +13 -3
- package/src/__tests__/sequence-execution.test.ts +22 -1
- package/src/__tests__/suno-credit-type.test.ts +21 -2
- package/src/__tests__/transcribe-capabilities.test.ts +41 -2
- package/src/__tests__/transcribe-preflight.test.ts +50 -1
- package/src/__tests__/video-analysis-catalog-sync.test.ts +4 -2
- package/src/__tests__/video-analysis-pricing.test.ts +10 -9
- package/src/__tests__/video-audio-capability.test.ts +35 -6
- package/src/__tests__/video-resolution-billed-as-named.test.ts +30 -0
- package/src/__tests__/video-upscale-params.test.ts +62 -0
- package/src/caption-styles.ts +62 -0
- package/src/catalog-projection.ts +8 -0
- package/src/credit-estimators/add-captions-nemotron.ts +73 -0
- package/src/credit-estimators/index.ts +20 -0
- package/src/credit-estimators/nemotron-transcribe.ts +87 -0
- package/src/credit-identifiers.ts +351 -41
- package/src/entries/captions.ts +2 -0
- package/src/entries/credits.ts +37 -2
- package/src/entries/llm-models.ts +6 -0
- package/src/entries/model-catalog.ts +22 -0
- package/src/entries/models.ts +50 -0
- package/src/entries/nodes.ts +4 -0
- package/src/entries/pipeline.ts +1 -1
- package/src/entries/video.ts +1 -1
- package/src/flux-3-klein.ts +69 -0
- package/src/freecut-protocol.ts +8 -0
- package/src/input-image-pricing.ts +23 -6
- package/src/lip-sync-levers.ts +56 -0
- package/src/lip-sync-pricing.ts +91 -1
- package/src/llm-models.ts +305 -16
- package/src/model-catalog.ts +529 -28
- package/src/model-constants.ts +417 -41
- package/src/model-tree.ts +2 -1
- package/src/nodaro-video-upscale.ts +48 -0
- package/src/node-default-mappings.ts +6 -0
- package/src/node-runtime-keys.ts +3 -0
- package/src/normalize-node-params.ts +172 -3
- package/src/runtime-config/contract.ts +10 -7
- package/src/seedance-draft.ts +2 -1
- package/src/seedance-extend-rate.ts +116 -0
- package/src/sequence-execution.ts +24 -0
- package/src/transcribe-preflight.ts +81 -7
- package/src/video-analysis-pricing.ts +28 -15
- package/src/video-duration.ts +9 -0
- package/src/video-ui-defaults.ts +5 -3
- package/src/video-upscale-params.ts +67 -0
- package/dist/chunk-2W745U4A.cjs.map +0 -1
- package/dist/chunk-3M4BNKFJ.cjs.map +0 -1
- package/dist/chunk-3OYO3XCC.js.map +0 -1
- package/dist/chunk-5YYN5X57.js.map +0 -1
- package/dist/chunk-6CSXMWL4.cjs.map +0 -1
- package/dist/chunk-6G7NN4S4.js.map +0 -1
- package/dist/chunk-AFYTX4AI.cjs.map +0 -1
- package/dist/chunk-AOJ6B7K4.cjs.map +0 -1
- package/dist/chunk-BLJUCPMV.cjs.map +0 -1
- package/dist/chunk-E2HZTGYX.js.map +0 -1
- package/dist/chunk-FVII3KHN.js.map +0 -1
- package/dist/chunk-GSFB537T.js.map +0 -1
- package/dist/chunk-HSHCCBIV.js.map +0 -1
- package/dist/chunk-IDTKF5EX.cjs.map +0 -1
- package/dist/chunk-IWWVPR3W.cjs.map +0 -1
- package/dist/chunk-KJSRPHJ2.cjs.map +0 -1
- package/dist/chunk-KYMHFVAZ.cjs.map +0 -1
- package/dist/chunk-MYJQELFN.cjs.map +0 -1
- package/dist/chunk-NGKKNSK6.js.map +0 -1
- package/dist/chunk-PTD6EDUU.js.map +0 -1
- package/dist/chunk-Q6IES3U2.js.map +0 -1
- package/dist/chunk-QWYFHQYK.cjs +0 -357
- package/dist/chunk-QWYFHQYK.cjs.map +0 -1
- package/dist/chunk-S24ZBUB4.js.map +0 -1
- package/dist/chunk-SYWR5G3T.js.map +0 -1
- package/dist/chunk-TRVWMKNE.cjs.map +0 -1
- package/dist/chunk-UDDVDQMF.js.map +0 -1
- package/dist/chunk-VACPZPUA.cjs.map +0 -1
- package/dist/chunk-WRSWH7US.cjs.map +0 -1
- package/dist/chunk-Y6X7UJSD.js.map +0 -1
- package/dist/chunk-YCXRSWCQ.js.map +0 -1
package/src/llm-models.ts
CHANGED
|
@@ -75,6 +75,38 @@ export interface LlmModelDef {
|
|
|
75
75
|
* Capability-driven so `llmCompleteStructured` never hardcodes provider ids.
|
|
76
76
|
*/
|
|
77
77
|
structuredOutputMode?: "anthropic-tool" | "kie-response-format" | "responses-json-schema"
|
|
78
|
+
/**
|
|
79
|
+
* For a model with NO trusted `structuredOutputMode`: the wire field a
|
|
80
|
+
* structured call still sends its schema in, as a hint the model reads but
|
|
81
|
+
* the provider was not found to hold its answer to. Read through
|
|
82
|
+
* {@link structuredWireMode}, which every request builder uses; everything
|
|
83
|
+
* that decides TRUST (STRUCTURED_VISION_MODELS, the structured-model
|
|
84
|
+
* filters) keeps reading `structuredOutputMode` alone, so a hinted model is
|
|
85
|
+
* never offered where guaranteed structured output is required. Every answer
|
|
86
|
+
* is still parsed, validated and re-asked by `llmCompleteStructured`, as for
|
|
87
|
+
* any model. Never set together with `structuredOutputMode` (registry test).
|
|
88
|
+
*/
|
|
89
|
+
structuredOutputHint?: "kie-response-format" | "responses-json-schema"
|
|
90
|
+
/**
|
|
91
|
+
* The KIE lane sends this model's structured calls NO `response_format`: the
|
|
92
|
+
* schema rides the system prompt instead (the one-object rule and the schema,
|
|
93
|
+
* as the KIE Claude lane's prompt-JSON), in the wire form `response_format`
|
|
94
|
+
* sent (keyword-named properties aliased and restored on the answer; Google's
|
|
95
|
+
* compact form), and `llmCompleteStructured` parses, validates and re-asks the
|
|
96
|
+
* answer as it does for any model. When every KIE answer of an unpinned call
|
|
97
|
+
* fails validation, the direct lane, which enforces the schema, is asked once
|
|
98
|
+
* before the call fails. For a model whose KIE endpoint fails a request that
|
|
99
|
+
* carries `response_format`.
|
|
100
|
+
*
|
|
101
|
+
* Per LANE, not per model: `structuredOutputMode` stays, so the trust readers
|
|
102
|
+
* (STRUCTURED_VISION_MODELS, the structured-model filters) and the direct
|
|
103
|
+
* lane's enforced `responseJsonSchema` are unchanged. Nothing reads trust to
|
|
104
|
+
* skip validation, and only an `anthropic-tool` model streams partial values,
|
|
105
|
+
* so a KIE-served answer here is checked like any untrusted one. Only on a
|
|
106
|
+
* `kieFormat: "chat-completions"` model whose wire mode is
|
|
107
|
+
* `kie-response-format` (registry test).
|
|
108
|
+
*/
|
|
109
|
+
kieSchemaInPrompt?: true
|
|
78
110
|
/** If set, fallback to direct Anthropic SDK with this model ID when KIE.ai fails */
|
|
79
111
|
directFallbackModel?: string
|
|
80
112
|
/**
|
|
@@ -118,6 +150,27 @@ export interface LlmModelDef {
|
|
|
118
150
|
directOnly?: true
|
|
119
151
|
/** Effort levels this model accepts (ascending). Absent/empty = no effort lever, picker hidden. */
|
|
120
152
|
reasoningEfforts?: readonly LlmReasoningEffort[]
|
|
153
|
+
/**
|
|
154
|
+
* true = a REQUESTED effort below this model's lowest level is sent as that
|
|
155
|
+
* lowest level, never as nothing ({@link effectiveReasoningEffort}): with no
|
|
156
|
+
* effort on the wire the provider runs this model at its HIGHEST level, so
|
|
157
|
+
* dropping a `none` request would buy the most reasoning instead of the
|
|
158
|
+
* least. Auto — no effort requested at all — still sends nothing. Absent =
|
|
159
|
+
* a request below the lowest level is dropped (Auto), the usual rule.
|
|
160
|
+
*/
|
|
161
|
+
effortBelowLowestSendsLowest?: true
|
|
162
|
+
/**
|
|
163
|
+
* true = the PRICE of a call on this model depends on whether it sends a
|
|
164
|
+
* reasoning effort, not only on its lane: the aggregator adds input of its
|
|
165
|
+
* own to every call, whose size depends on it. So the model's price key
|
|
166
|
+
* names the effort a call sends ({@link llmPriceKey}: `:effort-<level>`),
|
|
167
|
+
* and every estimate surface quotes that call. The sizes themselves are
|
|
168
|
+
* backend data (`backend/src/lib/pricing/llm-cost.ts`), held to this flag by
|
|
169
|
+
* a backend test. No model sets it today (GPT-6.1 Terra did for part of
|
|
170
|
+
* 2026-11-12, until its aggregator's added input stopped depending on the
|
|
171
|
+
* effort).
|
|
172
|
+
*/
|
|
173
|
+
effortPriced?: true
|
|
121
174
|
/**
|
|
122
175
|
* Effort levels available on the DIRECT lane, when the vendor's own API
|
|
123
176
|
* accepts more than the aggregator does. Absent = the direct lane offers the
|
|
@@ -317,6 +370,9 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
317
370
|
retired: true,
|
|
318
371
|
vendor: "google",
|
|
319
372
|
structuredOutputMode: "kie-response-format",
|
|
373
|
+
// The KIE endpoint is 3.8's, so it takes 3.8's rule: no `response_format`
|
|
374
|
+
// there, the schema in the prompt (see gemini-3.8-flash below).
|
|
375
|
+
kieSchemaInPrompt: true,
|
|
320
376
|
supportsImages: true,
|
|
321
377
|
// Google's own cap is 65,536, but the field feeds BOTH lanes and the KIE
|
|
322
378
|
// flash endpoints cap at 8192 (the measured 3.6 posture) — stay at the
|
|
@@ -339,8 +395,9 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
339
395
|
series: "Gemini",
|
|
340
396
|
// Inherits "Newest" from 3.7 (demoted above). Deliberately does NOT name
|
|
341
397
|
// agentic video: that capability exists only on the direct Google lane
|
|
342
|
-
// (Interactions API, not wired here)
|
|
343
|
-
//
|
|
398
|
+
// (Interactions API, not wired here). Plain video + audio input is wired
|
|
399
|
+
// (LLM_MODALITY_CAPS below) — a picker desc must not promise a lever this
|
|
400
|
+
// entry cannot expose.
|
|
344
401
|
desc: "Newest fast Gemini, agentic-tuned",
|
|
345
402
|
tier: "economy",
|
|
346
403
|
kieFormat: "chat-completions",
|
|
@@ -354,6 +411,16 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
354
411
|
// is ENFORCED (the reply came back exact-schema valid, credits_consumed
|
|
355
412
|
// present) — not merely accepted-and-ignored.
|
|
356
413
|
structuredOutputMode: "kie-response-format",
|
|
414
|
+
// Since 2026-11-13 the KIE endpoint answers ANY request carrying
|
|
415
|
+
// `response_format` (json_schema strict or not, json_object; any effort or
|
|
416
|
+
// none; with or without media) with HTTP 200 and an in-band
|
|
417
|
+
// `{"code":500}` after ~34 s, while the same request without it answers;
|
|
418
|
+
// 3.6 Flash, 3.1 Pro and Carbon take the same body. So the KIE lane sends
|
|
419
|
+
// the schema in the prompt (measured: a real Video Analysis roll, video
|
|
420
|
+
// included, answered schema-valid that way). The direct lane keeps its
|
|
421
|
+
// enforced `responseJsonSchema` (asked once when a call's KIE answers never
|
|
422
|
+
// validate), and the trust this mode carries stays.
|
|
423
|
+
kieSchemaInPrompt: true,
|
|
357
424
|
supportsImages: true,
|
|
358
425
|
// 16384 here, NOT the 8192 its 3.6 / 3.7 siblings sit at. Those two stay at
|
|
359
426
|
// the KIE-safe intersection because nobody measured their endpoints past
|
|
@@ -564,6 +631,54 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
564
631
|
// ended early (one declined the task, one stopped on RECITATION).
|
|
565
632
|
thinkingDefaultOn: true,
|
|
566
633
|
},
|
|
634
|
+
{
|
|
635
|
+
id: "gemini-4-carbon",
|
|
636
|
+
displayName: "Gemini 4 Carbon",
|
|
637
|
+
series: "Gemini",
|
|
638
|
+
desc: "Deepest-reasoning Gemini, 1M context",
|
|
639
|
+
tier: "premium",
|
|
640
|
+
kieFormat: "chat-completions",
|
|
641
|
+
// KIE serves it on the OpenAI-compatible dialect under this slug
|
|
642
|
+
// (docs.kie.ai/market/gemini/gemini-4-carbon-openai — Argon's page with the
|
|
643
|
+
// name swapped). Unlike Argon's, this lane passed its probe (2026-11-12):
|
|
644
|
+
// every call answered, images read, a stream ended cleanly, no empty 200,
|
|
645
|
+
// no negative counts, no code-execution fences. KIE flags it "Chat only",
|
|
646
|
+
// like Argon: it may not back a tool-loop feature (none does today).
|
|
647
|
+
kieSlugOrModel: "gemini-4-carbon-openai",
|
|
648
|
+
vendor: "google",
|
|
649
|
+
// NO trusted `structuredOutputMode` (decided 2026-11-12): KIE did NOT
|
|
650
|
+
// enforce a `response_format` json_schema — both probe calls came back as
|
|
651
|
+
// JSON holding an off-enum value in the enum field ("cheerful and
|
|
652
|
+
// energized" for happy | sad). So it is not a STRUCTURED_VISION_MODELS
|
|
653
|
+
// member (not offered by describe-to-picker), and every answer is parsed,
|
|
654
|
+
// validated and re-asked by `llmCompleteStructured`.
|
|
655
|
+
// The schema still goes on the wire, as a hint (`structuredOutputHint`): as
|
|
656
|
+
// `response_format` on KIE (the probe's replies followed its shape, only the
|
|
657
|
+
// enum slipped) and as `responseJsonSchema` on the direct Google lane,
|
|
658
|
+
// which enforced it in the probe — on every attempt, the re-ask included.
|
|
659
|
+
structuredOutputHint: "kie-response-format",
|
|
660
|
+
supportsImages: true,
|
|
661
|
+
// The registry's usual 16384, as Argon. Every call is floored to the
|
|
662
|
+
// default 32768 (`thinkingDefaultOn`, below): Google lists an output cap of
|
|
663
|
+
// 32,768 (models.list) and accepted maxOutputTokens 65536; KIE's doc
|
|
664
|
+
// states up to 65,536 tokens, thinking included.
|
|
665
|
+
maxOutputTokens: 16384,
|
|
666
|
+
// KIE first, Google as the reliability fallback (and Advanced mode) — no
|
|
667
|
+
// `preferDirect`, no `directOnly`. Google serves `models/gemini-4-carbon`
|
|
668
|
+
// beside Argon (models.list, 2026-11-12).
|
|
669
|
+
directGeminiModel: "gemini-4-carbon",
|
|
670
|
+
// KIE's documented `reasoning_effort` enum, and the direct lane's ladder
|
|
671
|
+
// too: Google 400s thinkingLevel MINIMAL on this model ("not supported")
|
|
672
|
+
// and answers LOW / MEDIUM / HIGH (probed 2026-11-12) — Argon's shape. So
|
|
673
|
+
// no `none` (it maps to MINIMAL in the direct client) and no
|
|
674
|
+
// `directReasoningEfforts`: a `none` request clamps to Auto on both lanes.
|
|
675
|
+
reasoningEfforts: ["low", "medium", "high"],
|
|
676
|
+
// Reasons with no thinking param sent on both lanes: KIE's `total_tokens`
|
|
677
|
+
// ran well past prompt + completion on every no-effort call (its dialect
|
|
678
|
+
// reports no reasoning split), and Google's no-level call reported thought
|
|
679
|
+
// tokens. No `reasoningOutputFloor`: both lanes take the default 32768.
|
|
680
|
+
thinkingDefaultOn: true,
|
|
681
|
+
},
|
|
567
682
|
{
|
|
568
683
|
id: "claude-opus-4.7",
|
|
569
684
|
displayName: "Claude Opus 4.7",
|
|
@@ -807,6 +922,43 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
807
922
|
supportsTemperature: false,
|
|
808
923
|
thinkingDefaultOn: true,
|
|
809
924
|
},
|
|
925
|
+
{
|
|
926
|
+
id: "gpt-6.1-terra",
|
|
927
|
+
displayName: "GPT-6.1 Terra",
|
|
928
|
+
series: "GPT",
|
|
929
|
+
desc: "Fast, efficient GPT-6.1, sharp visual reasoning",
|
|
930
|
+
tier: "standard",
|
|
931
|
+
kieFormat: "responses",
|
|
932
|
+
// The GPT-6 wire (codex/v1/responses, derived from `vendor`); KIE's doc is
|
|
933
|
+
// 6.1 Sol's page with the name swapped and `none` added to the effort enum.
|
|
934
|
+
// Live-probed 2026-11-12: non-stream 16/16, `text.format` json_schema
|
|
935
|
+
// ENFORCED (asked for prose, it returned schema JSON with the enum
|
|
936
|
+
// honoured), data-URI vision, SSE deltas. Unlike luna / sol / 6.1 sol it
|
|
937
|
+
// reports REAL top-level `usage.input_tokens` — see `kieInputUntrusted` in
|
|
938
|
+
// backend/src/lib/llm/kie-cost.ts.
|
|
939
|
+
//
|
|
940
|
+
// KIE's endpoint adds its own instructions to every call and counts them as
|
|
941
|
+
// input. The hold and the estimate count them too
|
|
942
|
+
// (backend/src/lib/pricing/llm-cost.ts, `llmLaneAddedInputTokens`). They
|
|
943
|
+
// were several times larger with an effort sent on the morning of
|
|
944
|
+
// 2026-11-12 and the same size either way by that afternoon, so its price
|
|
945
|
+
// does not depend on the effort: not `effortPriced` (re-measure if KIE's
|
|
946
|
+
// reported input jumps again).
|
|
947
|
+
kieSlugOrModel: "gpt-6-1-terra",
|
|
948
|
+
vendor: "openai",
|
|
949
|
+
structuredOutputMode: "responses-json-schema",
|
|
950
|
+
supportsImages: true,
|
|
951
|
+
maxOutputTokens: 16384,
|
|
952
|
+
// NO `none` (decided 2026-11-12): KIE answers it, but a `none` sent on the
|
|
953
|
+
// wire is answered as an effort (it echoed `medium`) while reasoning no
|
|
954
|
+
// more than Auto. Clamping maps a `none` request to nothing sent (Auto).
|
|
955
|
+
reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
|
|
956
|
+
// Same doc as 6.1 Sol: no temperature lever.
|
|
957
|
+
supportsTemperature: false,
|
|
958
|
+
// NO `thinkingDefaultOn`: with no effort sent (and at low / medium / high)
|
|
959
|
+
// the probe reported 0 reasoning tokens; only xhigh (8) and max (27)
|
|
960
|
+
// reasoned, and those already floor the cap (`raiseToReasoningFloor`).
|
|
961
|
+
},
|
|
810
962
|
{
|
|
811
963
|
id: "grok-4.6",
|
|
812
964
|
displayName: "Grok 4.6",
|
|
@@ -1036,12 +1188,49 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
1036
1188
|
// KIE's documented enum. The endpoint accepted xhigh/none without error but
|
|
1037
1189
|
// never echoes an effort back, so nothing beyond the doc is verifiable.
|
|
1038
1190
|
reasoningEfforts: ["low", "medium", "high"],
|
|
1191
|
+
// Decided 2026-11-12: a request below `low` (`none`) is sent as `low`,
|
|
1192
|
+
// never as nothing — KIE runs Kimi at its highest level when no effort is
|
|
1193
|
+
// sent (K3.1 Pro's no-effort replies echoed `max`). Auto still sends nothing.
|
|
1194
|
+
effortBelowLowestSendsLowest: true,
|
|
1039
1195
|
// The doc lists no `temperature` field at all.
|
|
1040
1196
|
supportsTemperature: false,
|
|
1041
1197
|
// "Reasons by default" per KIE's doc (its usage reports no reasoning-token
|
|
1042
1198
|
// split, so this is the doc's word, not a measurement).
|
|
1043
1199
|
thinkingDefaultOn: true,
|
|
1044
1200
|
},
|
|
1201
|
+
{
|
|
1202
|
+
id: "kimi-k3.1-pro",
|
|
1203
|
+
displayName: "Kimi K3.1 Pro",
|
|
1204
|
+
series: "Kimi",
|
|
1205
|
+
desc: "Kimi tuned for coding and research, 1M context",
|
|
1206
|
+
tier: "premium",
|
|
1207
|
+
kieFormat: "responses",
|
|
1208
|
+
// Same openai/v1/responses family path and wire as kimi-k3 (KIE's doc is
|
|
1209
|
+
// K3's page with the name and the effort enum changed). Live-probed
|
|
1210
|
+
// 2026-11-12: non-stream 9/9, `text.format` json_schema ENFORCED (asked for
|
|
1211
|
+
// prose, it returned schema JSON with the enum honoured), data-URI vision,
|
|
1212
|
+
// real `usage.input_tokens` and `credits_consumed`. Its stream sends
|
|
1213
|
+
// `response.reasoning_text.delta` frames apart from the
|
|
1214
|
+
// `response.output_text.delta` ones, and only the latter are the answer.
|
|
1215
|
+
kieSlugOrModel: "kimi-k3-1-pro",
|
|
1216
|
+
vendor: "moonshot",
|
|
1217
|
+
structuredOutputMode: "responses-json-schema",
|
|
1218
|
+
supportsImages: true,
|
|
1219
|
+
maxOutputTokens: 16384,
|
|
1220
|
+
// KIE's documented enum (K3's was low / medium / high). Each level was
|
|
1221
|
+
// echoed back. An off-enum `medium` answered too (echoed as `high`), so it
|
|
1222
|
+
// is not offered: a `medium` request clamps down to `low`.
|
|
1223
|
+
reasoningEfforts: ["low", "high", "max"],
|
|
1224
|
+
// Decided 2026-11-12: a request below `low` (`none`) is sent as `low`,
|
|
1225
|
+
// never as nothing — with no effort sent KIE runs it at `max` (the probe's
|
|
1226
|
+
// no-effort replies echoed `max`). Auto still sends nothing.
|
|
1227
|
+
effortBelowLowestSendsLowest: true,
|
|
1228
|
+
// The doc lists no `temperature` field at all, as K3's.
|
|
1229
|
+
supportsTemperature: false,
|
|
1230
|
+
// Measured: every probe call reported reasoning tokens (20 to 163), the
|
|
1231
|
+
// no-effort ones included (they echoed effort `max`).
|
|
1232
|
+
thinkingDefaultOn: true,
|
|
1233
|
+
},
|
|
1045
1234
|
{
|
|
1046
1235
|
id: "deepseek-v4.1-flash",
|
|
1047
1236
|
displayName: "DeepSeek V4.1 Flash",
|
|
@@ -1099,6 +1288,33 @@ export const STRUCTURED_VISION_MODELS = LLM_MODELS.filter(
|
|
|
1099
1288
|
(m) => m.supportsImages && m.structuredOutputMode != null,
|
|
1100
1289
|
)
|
|
1101
1290
|
|
|
1291
|
+
/**
|
|
1292
|
+
* How a structured call's schema goes on the wire for `model`: its trusted
|
|
1293
|
+
* `structuredOutputMode`, else its `structuredOutputHint`; undefined = the
|
|
1294
|
+
* schema is not sent at all (a model with neither: plain text, parsed and
|
|
1295
|
+
* re-asked). What the request builders read — never a TRUST decision, which
|
|
1296
|
+
* reads `structuredOutputMode` alone.
|
|
1297
|
+
*/
|
|
1298
|
+
export function structuredWireMode(
|
|
1299
|
+
model: Pick<LlmModelDef, "structuredOutputMode" | "structuredOutputHint">,
|
|
1300
|
+
): LlmModelDef["structuredOutputMode"] {
|
|
1301
|
+
return model.structuredOutputMode ?? model.structuredOutputHint
|
|
1302
|
+
}
|
|
1303
|
+
|
|
1304
|
+
/**
|
|
1305
|
+
* Does a structured call on `modelId` (unset = `feature`'s default model) send
|
|
1306
|
+
* its schema ({@link structuredWireMode})? On the wire, or in the system prompt
|
|
1307
|
+
* on a lane that carries it there (`kieSchemaInPrompt`): either way the model
|
|
1308
|
+
* reads it, so a request's size counts it. A schema that is not sent is not
|
|
1309
|
+
* read by the model. A model the registry does not know counts as yes (the
|
|
1310
|
+
* larger size).
|
|
1311
|
+
*/
|
|
1312
|
+
export function llmSendsSchema(modelId: string | undefined, feature?: string): boolean {
|
|
1313
|
+
const id = modelId || (feature ? LLM_FEATURE_DEFAULTS[feature as LlmFeature] : undefined)
|
|
1314
|
+
const model = id ? getLlmModel(id) : undefined
|
|
1315
|
+
return !model || structuredWireMode(model) !== undefined
|
|
1316
|
+
}
|
|
1317
|
+
|
|
1102
1318
|
/**
|
|
1103
1319
|
* Models {@link STRUCTURED_VISION_MODELS} carries that describe-to-picker does
|
|
1104
1320
|
* NOT offer — now none.
|
|
@@ -1352,6 +1568,10 @@ export function defaultReasoningEffort(feature: string, modelId?: string): LlmRe
|
|
|
1352
1568
|
* input requires a separate audio-capable model we don't route to today.
|
|
1353
1569
|
*/
|
|
1354
1570
|
export const LLM_MODALITY_CAPS: Record<string, { image: boolean; video: boolean; audio: boolean }> = {
|
|
1571
|
+
// Keeps full caps after KIE dropped its channel (2026-11-13): it stays a
|
|
1572
|
+
// VIDEO_ANALYSIS_LLM_MODELS member as the legacy `fast` backing (below), so a
|
|
1573
|
+
// stored raw `llmModel: "gemini-3-flash"` still resolves and prices under its
|
|
1574
|
+
// own credit family.
|
|
1355
1575
|
"gemini-3-flash": { image: true, video: true, audio: true },
|
|
1356
1576
|
"gemini-3.6-flash": { image: true, video: true, audio: true },
|
|
1357
1577
|
// gemini-3.7-flash is IMAGE-ONLY by DECISION, not omission: full video+audio
|
|
@@ -1360,15 +1580,13 @@ export const LLM_MODALITY_CAPS: Record<string, { image: boolean; video: boolean;
|
|
|
1360
1580
|
// deferred while the smart-family A/B routes this model internally (#747).
|
|
1361
1581
|
// Flip these two flags ONLY together with that VA-side decision.
|
|
1362
1582
|
"gemini-3.7-flash": { image: true, video: false, audio: false },
|
|
1363
|
-
// gemini-3.8-flash
|
|
1364
|
-
//
|
|
1365
|
-
//
|
|
1366
|
-
//
|
|
1367
|
-
//
|
|
1368
|
-
//
|
|
1369
|
-
|
|
1370
|
-
// with that VA-side decision.
|
|
1371
|
-
"gemini-3.8-flash": { image: true, video: false, audio: false },
|
|
1583
|
+
// gemini-3.8-flash takes video + audio (decided 2026-11-13): it backs the
|
|
1584
|
+
// video-analysis `fast` tier (VIDEO_ANALYSIS_TIERS below) since KIE dropped
|
|
1585
|
+
// gemini-3-flash's channel, which is what enrols it in
|
|
1586
|
+
// VIDEO_ANALYSIS_LLM_MODELS. Its KIE lane reads a video URL sent as an
|
|
1587
|
+
// `image_url` part, the same channel gemini-3-flash used (live-verified
|
|
1588
|
+
// 2026-11-13 with the analysis request: prompt tokens include the clip).
|
|
1589
|
+
"gemini-3.8-flash": { image: true, video: true, audio: true },
|
|
1372
1590
|
"gemini-3.1-pro": { image: true, video: true, audio: true },
|
|
1373
1591
|
// gemini-4-argon is IMAGE-ONLY by DECISION, not omission (decided 2026-10-17):
|
|
1374
1592
|
// the model DOES read video and audio, but full caps would
|
|
@@ -1376,6 +1594,12 @@ export const LLM_MODALITY_CAPS: Record<string, { image: boolean; video: boolean;
|
|
|
1376
1594
|
// video-analysis tier + pricing decision. That waits on a Video Analysis
|
|
1377
1595
|
// bake-off; flip these two flags ONLY together with its outcome.
|
|
1378
1596
|
"gemini-4-argon": { image: true, video: false, audio: false },
|
|
1597
|
+
// gemini-4-carbon is IMAGE-ONLY for now, by decision (2026-11-12): the model
|
|
1598
|
+
// reads video and audio, but full caps would auto-enroll it in
|
|
1599
|
+
// VIDEO_ANALYSIS_LLM_MODELS (derived below). Its video + audio input comes in
|
|
1600
|
+
// a follow-up change that also decides its Video Analysis enrolment; flip
|
|
1601
|
+
// these two flags ONLY together with that.
|
|
1602
|
+
"gemini-4-carbon": { image: true, video: false, audio: false },
|
|
1379
1603
|
"claude-haiku-4.5": { image: true, video: false, audio: false },
|
|
1380
1604
|
"claude-haiku-5.5": { image: true, video: false, audio: false },
|
|
1381
1605
|
"claude-sonnet-4.6": { image: true, video: false, audio: false },
|
|
@@ -1392,6 +1616,8 @@ export const LLM_MODALITY_CAPS: Record<string, { image: boolean; video: boolean;
|
|
|
1392
1616
|
"gpt-6-luna": { image: true, video: false, audio: false },
|
|
1393
1617
|
"gpt-6-sol": { image: true, video: false, audio: false },
|
|
1394
1618
|
"gpt-6.1-sol": { image: true, video: false, audio: false },
|
|
1619
|
+
// KIE's doc also takes files (PDFs, documents); that input is not wired here.
|
|
1620
|
+
"gpt-6.1-terra": { image: true, video: false, audio: false },
|
|
1395
1621
|
"grok-4.6": { image: true, video: false, audio: false },
|
|
1396
1622
|
"grok-4.7": { image: true, video: false, audio: false },
|
|
1397
1623
|
"claude-sonnet-5": { image: true, video: false, audio: false },
|
|
@@ -1402,6 +1628,7 @@ export const LLM_MODALITY_CAPS: Record<string, { image: boolean; video: boolean;
|
|
|
1402
1628
|
"claude-fable-5": { image: true, video: false, audio: false },
|
|
1403
1629
|
// KIE documents text + image input only for both.
|
|
1404
1630
|
"kimi-k3": { image: true, video: false, audio: false },
|
|
1631
|
+
"kimi-k3.1-pro": { image: true, video: false, audio: false },
|
|
1405
1632
|
"deepseek-v4.1-flash": { image: true, video: false, audio: false },
|
|
1406
1633
|
}
|
|
1407
1634
|
|
|
@@ -1479,7 +1706,12 @@ export function availableReasoningEfforts(
|
|
|
1479
1706
|
return model.reasoningEfforts ?? []
|
|
1480
1707
|
}
|
|
1481
1708
|
|
|
1482
|
-
/**
|
|
1709
|
+
/**
|
|
1710
|
+
* Highest level the model supports that is ≤ the requested level; undefined =
|
|
1711
|
+
* treat as Auto. A request below the model's lowest level is dropped (Auto) —
|
|
1712
|
+
* except on a model that declares `effortBelowLowestSendsLowest`, where it is
|
|
1713
|
+
* sent as that lowest level. No request at all is always Auto.
|
|
1714
|
+
*/
|
|
1483
1715
|
export function effectiveReasoningEffort(
|
|
1484
1716
|
modelId: string | undefined,
|
|
1485
1717
|
requested?: string,
|
|
@@ -1493,6 +1725,7 @@ export function effectiveReasoningEffort(
|
|
|
1493
1725
|
for (const l of levels) {
|
|
1494
1726
|
if (EFFORT_RANK[l] <= EFFORT_RANK[req] && (best === undefined || EFFORT_RANK[l] > EFFORT_RANK[best])) best = l
|
|
1495
1727
|
}
|
|
1728
|
+
if (best === undefined && getLlmModel(modelId ?? "")?.effortBelowLowestSendsLowest) return levels[0]
|
|
1496
1729
|
return best
|
|
1497
1730
|
}
|
|
1498
1731
|
|
|
@@ -1610,12 +1843,16 @@ export function llmServesDirect(
|
|
|
1610
1843
|
* `POST /v1/credits/model-costs` (or `GET /v1/credits/model-cost?model=`) for
|
|
1611
1844
|
* the call's estimated price: `<feature>:<modelId>`, plus `:direct` when the
|
|
1612
1845
|
* call is served on the vendor's own API ({@link llmServesDirect}), which is
|
|
1613
|
-
* priced at the direct rate
|
|
1846
|
+
* priced at the direct rate, or — on a model whose price depends on the effort
|
|
1847
|
+
* (`effortPriced`, served on the aggregator) — `:effort-<level>` naming the
|
|
1848
|
+
* effort the call sends. On every other model the effort changes the key only
|
|
1849
|
+
* through its lane. Read back by {@link parseLlmPriceKeyModel}.
|
|
1614
1850
|
*
|
|
1615
1851
|
* With no `reasoningEffort` the default effort of the feature the call
|
|
1616
1852
|
* borrows its defaults from applies ({@link llmDefaultsFeature},
|
|
1617
1853
|
* {@link defaultReasoningEffort}) — the effort the call then runs at — so the
|
|
1618
|
-
* key names the lane the call is held and
|
|
1854
|
+
* key names the lane (and, where it matters, the effort) the call is held and
|
|
1855
|
+
* charged on.
|
|
1619
1856
|
*
|
|
1620
1857
|
* `modelId` must already be resolved: the model the call runs on, never empty
|
|
1621
1858
|
* (a caller with no model passes the feature's default, `LLM_FEATURE_DEFAULTS`
|
|
@@ -1629,7 +1866,53 @@ export function llmPriceKey(
|
|
|
1629
1866
|
): string {
|
|
1630
1867
|
if (!modelId) throw new TypeError(`llmPriceKey: modelId must be a resolved model id (feature "${feature}")`)
|
|
1631
1868
|
const effort = reasoningEffort ?? defaultReasoningEffort(llmDefaultsFeature(feature), modelId)
|
|
1632
|
-
|
|
1869
|
+
if (llmServesDirect(modelId, effort, advancedMode)) return `${feature}:${modelId}${LLM_PRICE_KEY_DIRECT}`
|
|
1870
|
+
// A model whose price depends on the effort (`effortPriced`) names the one
|
|
1871
|
+
// the call sends — clamped as the aggregator's wire clamps it, so a level it
|
|
1872
|
+
// does not offer is the effort that is really sent (or none). Every other
|
|
1873
|
+
// model's key is unchanged.
|
|
1874
|
+
const sent = getLlmModel(modelId)?.effortPriced ? effectiveReasoningEffort(modelId, effort) : undefined
|
|
1875
|
+
return sent ? `${feature}:${modelId}${LLM_PRICE_KEY_EFFORT}${sent}` : `${feature}:${modelId}`
|
|
1876
|
+
}
|
|
1877
|
+
|
|
1878
|
+
/** The price-key suffix of a call served on the vendor's own API (see {@link llmPriceKey}). */
|
|
1879
|
+
export const LLM_PRICE_KEY_DIRECT = ":direct"
|
|
1880
|
+
/** The price-key suffix naming the effort a call sends, on an `effortPriced` model: `:effort-<level>`. */
|
|
1881
|
+
export const LLM_PRICE_KEY_EFFORT = ":effort-"
|
|
1882
|
+
|
|
1883
|
+
/** A price key's model part, read back by {@link parseLlmPriceKeyModel}. */
|
|
1884
|
+
export interface LlmPriceKeyModel {
|
|
1885
|
+
modelId: string
|
|
1886
|
+
/** Present only for a key that names the direct lane. */
|
|
1887
|
+
lane?: "direct"
|
|
1888
|
+
/** Present only for a key that names an effort (an `effortPriced` model's). */
|
|
1889
|
+
reasoningEffort?: LlmReasoningEffort
|
|
1890
|
+
}
|
|
1891
|
+
|
|
1892
|
+
const REGISTRY_MODEL_IDS: ReadonlySet<string> = new Set(LLM_MODEL_IDS)
|
|
1893
|
+
|
|
1894
|
+
/**
|
|
1895
|
+
* The part of a price key after its `<feature>:` — `<modelId>`,
|
|
1896
|
+
* `<modelId>:direct` or `<modelId>:effort-<level>` — read back, or null for
|
|
1897
|
+
* anything that is not a key {@link llmPriceKey} could build: a model that is
|
|
1898
|
+
* not a registry id (never a borrowed slug), an effort on a model whose price
|
|
1899
|
+
* does not depend on it, a level it does not offer on the aggregator, any
|
|
1900
|
+
* other suffix. The one reader both the server and the editor use, so every
|
|
1901
|
+
* key names exactly one price.
|
|
1902
|
+
*/
|
|
1903
|
+
export function parseLlmPriceKeyModel(tail: string): LlmPriceKeyModel | null {
|
|
1904
|
+
if (REGISTRY_MODEL_IDS.has(tail)) return { modelId: tail }
|
|
1905
|
+
if (tail.endsWith(LLM_PRICE_KEY_DIRECT)) {
|
|
1906
|
+
const modelId = tail.slice(0, -LLM_PRICE_KEY_DIRECT.length)
|
|
1907
|
+
return REGISTRY_MODEL_IDS.has(modelId) ? { modelId, lane: "direct" } : null
|
|
1908
|
+
}
|
|
1909
|
+
const at = tail.lastIndexOf(LLM_PRICE_KEY_EFFORT)
|
|
1910
|
+
if (at <= 0) return null
|
|
1911
|
+
const modelId = tail.slice(0, at)
|
|
1912
|
+
const level = tail.slice(at + LLM_PRICE_KEY_EFFORT.length) as LlmReasoningEffort
|
|
1913
|
+
const model = REGISTRY_MODEL_IDS.has(modelId) ? getLlmModel(modelId) : undefined
|
|
1914
|
+
if (!model?.effortPriced || !availableReasoningEfforts(modelId).includes(level)) return null
|
|
1915
|
+
return { modelId, reasoningEffort: level }
|
|
1633
1916
|
}
|
|
1634
1917
|
|
|
1635
1918
|
/**
|
|
@@ -1715,7 +1998,7 @@ export const VIDEO_ANALYSIS_LLM_MODELS: string[] = LLM_MODELS
|
|
|
1715
1998
|
* VIDEO_ANALYSIS_LLM_MODELS member AND every such model is reachable by a tier,
|
|
1716
1999
|
* so adding a video model forces a tier decision instead of silently leaking.
|
|
1717
2000
|
*/
|
|
1718
|
-
export const VIDEO_ANALYSIS_TIERS = { fast: "gemini-3-flash", pro: "gemini-3.1-pro" } as const
|
|
2001
|
+
export const VIDEO_ANALYSIS_TIERS = { fast: "gemini-3.8-flash", pro: "gemini-3.1-pro" } as const
|
|
1719
2002
|
export type VideoAnalysisModelTier = keyof typeof VIDEO_ANALYSIS_TIERS
|
|
1720
2003
|
/**
|
|
1721
2004
|
* Models that PREVIOUSLY backed a tier, mapped to the tier they backed.
|
|
@@ -1736,6 +2019,12 @@ export const VIDEO_ANALYSIS_LEGACY_MODELS: Record<string, VideoAnalysisModelTier
|
|
|
1736
2019
|
// belongs here: stored raw `llmModel` values keep resolving and keep pricing
|
|
1737
2020
|
// under their own credit family.
|
|
1738
2021
|
"gemini-3.6-flash": "fast",
|
|
2022
|
+
// Backed the fast tier from 2026-07-29 until 2026-11-13, when KIE dropped its
|
|
2023
|
+
// channel (every call answers `code 422 "The channel is not supported"`) and
|
|
2024
|
+
// the fast tier moved to gemini-3.8-flash. Stored raw `llmModel` values keep
|
|
2025
|
+
// resolving and keep pricing under their own credit family; the analysis
|
|
2026
|
+
// engine serves their rolls with the current fast model.
|
|
2027
|
+
"gemini-3-flash": "fast",
|
|
1739
2028
|
}
|
|
1740
2029
|
/**
|
|
1741
2030
|
* MIXED tiers — advanced multi-engine analysis plans whose identifier resolves
|