@grida/ai-models 0.0.0-stage → 1.20261011.1

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/LICENSE ADDED
@@ -0,0 +1,21 @@
1
+ MIT License
2
+
3
+ Copyright (c) 2026 Grida
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
package/PROVENANCE.md ADDED
@@ -0,0 +1,651 @@
1
+ # Catalogue research notes
2
+
3
+ Source-specific qualifications and references for the values in
4
+ [`models.json`](./models.json). The JSON holds the values; these notes explain
5
+ where they came from and any caveat a value cannot carry itself. Update a note
6
+ in the same change as the fact it qualifies.
7
+
8
+ Each `##` heading names the entry or declaration a note qualifies. The headings
9
+ are code fragments carried over from the TypeScript literals the catalogue was
10
+ first written as in [gridaco/grida](https://github.com/gridaco/grida), so they
11
+ keep that historical spelling (for example `imageInputMimes` rather than
12
+ `image_input_mimes`).
13
+
14
+ ## `const OPENAI_IMAGE_INPUT_MIMES = [`
15
+
16
+ Provider-family capabilities are kept private so catalogue entries remain
17
+ explicit while sharing one source-backed value. Do not derive these from
18
+ `multimodal`; a future model may be multimodal without a documented image
19
+ input format set.
20
+ https://developers.openai.com/api/docs/guides/images-vision#image-input-requirements
21
+
22
+ ## `const ANTHROPIC_IMAGE_INPUT_MIMES = [`
23
+
24
+ https://platform.claude.com/docs/en/build-with-claude/vision#supported-formats
25
+
26
+ ## `const GOOGLE_IMAGE_INPUT_MIMES = [`
27
+
28
+ https://ai.google.dev/gemini-api/docs/image-understanding#supported-image-formats
29
+
30
+ ## `const OPENAI_LONG_CONTEXT_PRICING = {`
31
+
32
+ OpenAI bills the full request at these multipliers once its total input
33
+ exceeds 272K tokens. The same rule is published for GPT-5.5, every
34
+ GPT-5.6 family member, and GPT-6 Astra.
35
+ https://developers.openai.com/api/docs/models/gpt-5.5
36
+ https://developers.openai.com/api/docs/models/gpt-5.6-sol
37
+ https://developers.openai.com/api/docs/models/gpt-6-astra
38
+
39
+ ## `"openai/gpt-5.6-sol": {`
40
+
41
+ Base rates; OPENAI_LONG_CONTEXT_PRICING represents the request-wide
42
+ band that applies above 272K total input tokens.
43
+
44
+ Sol is cheaper than GPT-5.5, the card directly above it. It was
45
+ introduced carrying 5.5's rates verbatim; these are OpenAI's own.
46
+ https://developers.openai.com/api/docs/models/gpt-5.6-sol
47
+
48
+ ## `"openai/gpt-6-astra": {`
49
+
50
+ OpenAI's September 3 introduction remained a limited rollout. The
51
+ September 4 date below is the exact Vercel AI Gateway route's broad availability,
52
+ which is the release fact relevant to this route card.
53
+ https://openai.com/products/release-notes/
54
+ https://vercel.com/ai-gateway/models/gpt-6-astra
55
+
56
+ ## `"anthropic/claude-sonnet-5": {`
57
+
58
+ $2/$10 is the standard rate, not a live discount: it launched as an
59
+ introductory rate and Anthropic made it permanent, cancelling the
60
+ announced step up to $3/$15. Do not restore the higher card.
61
+ https://platform.claude.com/docs/en/about-claude/pricing
62
+
63
+ ## `"anthropic/claude-fable-5.1": {`
64
+
65
+ Same input/output/cacheWrite card as Claude Fable 5, with cache reads
66
+ cut to $0.25/MTok. Not a drop-in successor — forced tool choice
67
+ (`tool_choice` `any`/`tool`) is rejected here — which is why Fable 5
68
+ stays catalogued rather than being removed.
69
+ https://platform.claude.com/docs/en/models/fable-5-1/overview
70
+
71
+ ## `"anthropic/claude-fable-5": {`
72
+
73
+ Unlike Fable 5.1, this version accepts forced tool choice.
74
+
75
+ ## `"anthropic/claude-opus-5": {`
76
+
77
+ Drop-in successor to Opus 4.8 at the same rate card.
78
+
79
+ ## `"google/gemini-3.8-flash": {`
80
+
81
+ Google's cache model is read + hourly storage (no one-time write
82
+ premium that matches `cacheWrite` semantics), so the field is omitted.
83
+
84
+ Steady-state rates. Google is promoting $0.75 in / $3.75 out / $0.075
85
+ cacheRead through 2026-12-31; these are the prices that apply from
86
+ 2027-01-01. Checking Google's page before then will show the lower
87
+ set — that is the promotion, not a correction. Gemini 3.8 Flash is GA,
88
+ and this exact id is live on both Vercel AI Gateway and OpenRouter.
89
+ https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash
90
+ https://ai.google.dev/gemini-api/docs/pricing
91
+ https://vercel.com/ai-gateway/models/gemini-3.8-flash
92
+ https://openrouter.ai/google/gemini-3.8-flash
93
+
94
+ ## `"google/gemini-3.7-flash": {`
95
+
96
+ 3.8 improves accuracy and reliability at the same rate, but Google
97
+ still recommends 3.7 when compute efficiency matters because 3.8 can
98
+ consume more tokens.
99
+ https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/guides/gemini-3-8-flash
100
+
101
+ ## `longContext: {`
102
+
103
+ Google bills the full request at $4 in / $18 out / $0.40
104
+ cacheRead above 200K tokens. `inputMultiplier` covers every
105
+ input bucket, so cacheRead 0.2 x 2 lands on $0.40 without a
106
+ separate field.
107
+ https://ai.google.dev/gemini-api/docs/pricing
108
+
109
+ ## `const GPT_IMAGE_2_5_FAL_PRICING: PerTokenPricing = {`
110
+
111
+ Both 2.5 variants use this fal meter. Keep the separate text-output
112
+ rate: fal publishes it even though its result schema exposes only images.
113
+ https://fal.ai/models/openai/gpt-image-2.5/flare/text-to-image
114
+ https://fal.ai/models/openai/gpt-image-2.5/sunburst/text-to-image
115
+
116
+ ## `const GPT_IMAGE_2_5_VERCEL_AI_GATEWAY_PRICING: PerTokenPricing = {`
117
+
118
+ Each serving provider publishes its own meter; do not copy fal's extra
119
+ image-cache/text-output rates into providers that do not advertise them.
120
+ https://ai-gateway.vercel.sh/v1/models (verified 2026-09-09 KST)
121
+
122
+ ## `const GPT_IMAGE_2_5_OPENROUTER_PRICING: PerTokenPricing = {`
123
+
124
+ Both variants' /api/v1/images/models/{id}/endpoints publish this meter.
125
+ https://openrouter.ai/api/v1/images/models/openai/gpt-image-2.5-flare/endpoints
126
+ https://openrouter.ai/api/v1/images/models/openai/gpt-image-2.5-sunburst/endpoints
127
+
128
+ ## `"openai/gpt-image-2": {`
129
+
130
+ ---
131
+
132
+ OpenAI
133
+ -----------------------------------------------------------------
134
+
135
+ https://developers.openai.com/api/docs/models/gpt-image-2
136
+
137
+ ## `transparent_background: true,`
138
+
139
+ Native transparency added 2026-08-20; provider exposure differs.
140
+ https://developers.openai.com/api/docs/changelog
141
+
142
+ ## `providers: {`
143
+
144
+ ids/prices verified 2026-06-29, see github.com/gridaco/grida/issues/908
145
+
146
+ ## `transparent_background: null,`
147
+
148
+ This provider's page does not establish background support.
149
+ https://vercel.com/ai-gateway/models/gpt-image-2 (2026-09-09)
150
+
151
+ ## `transparent_background: false,`
152
+
153
+ Published background enum is auto | opaque (2026-09-09).
154
+ https://openrouter.ai/api/v1/images/models/openai/gpt-image-2/endpoints
155
+
156
+ ## `references: { id: "openai/gpt-image-2", max: 16 },`
157
+
158
+ input_references advertised by OpenRouter (0–16), 2026-07-01.
159
+
160
+ ## `},`
161
+
162
+ Inherits native transparency: fal-ai/gpt-image-2's published
163
+ background enum includes transparent (verified 2026-09-09).
164
+
165
+ ## `"openai/gpt-image-2.5-flare": {`
166
+
167
+ Released as two distinct models. All three providers now list both
168
+ variants; fal separates generation and edit endpoint ids, while
169
+ OpenRouter advertises input_references on the same id (2026-09-09 KST).
170
+
171
+ ## `transparent_background: true,`
172
+
173
+ Both fal generation and edit schemas expose background=transparent.
174
+ https://fal.ai/models/openai/gpt-image-2.5/flare/text-to-image/api
175
+ https://fal.ai/models/openai/gpt-image-2.5/flare/edit/api
176
+
177
+ ## `url: "https://vercel.com/ai-gateway/models/gpt-image-2.5-flare",`
178
+
179
+ The model page exposes background=transparent; reference input
180
+ is not established by the Vercel AI Gateway feed (2026-09-09 KST).
181
+
182
+ ## `transparent_background: false,`
183
+
184
+ Endpoint schema's background enum is auto | opaque.
185
+
186
+ ## `constraints: {`
187
+
188
+ Model-specific fal documentation overrides its generic ImageSize
189
+ component's wider bounds. The API also accepts auto-sized output.
190
+ https://fal.ai/models/openai/gpt-image-2.5/flare/text-to-image/api
191
+
192
+ ## `avg_cost_usd: 0.055,`
193
+
194
+ High 1024² output estimate is $0.05268, plus a small input allowance.
195
+ Not a fixed per-image price: actual cost depends on all billed tokens.
196
+ https://developers.openai.com/api/docs/guides/image-generation
197
+
198
+ ## `transparent_background: true,`
199
+
200
+ Both fal generation and edit schemas expose background=transparent.
201
+ https://fal.ai/models/openai/gpt-image-2.5/sunburst/text-to-image/api
202
+ https://fal.ai/models/openai/gpt-image-2.5/sunburst/edit/api
203
+
204
+ ## `url: "https://vercel.com/ai-gateway/models/gpt-image-2.5-sunburst",`
205
+
206
+ The model page exposes background=transparent; reference input
207
+ is not established by the Vercel AI Gateway feed (2026-09-09 KST).
208
+
209
+ ## `transparent_background: false,`
210
+
211
+ Endpoint schema's background enum is auto | opaque.
212
+
213
+ ## `constraints: {`
214
+
215
+ https://fal.ai/models/openai/gpt-image-2.5/sunburst/text-to-image/api
216
+
217
+ ## `"openai/gpt-image-1.5": {`
218
+
219
+ https://developers.openai.com/api/docs/models/gpt-image-1.5
220
+
221
+ ## `constraints: null,`
222
+
223
+ Preset-only — provider rejects arbitrary sizes.
224
+
225
+ ## `pricing: {`
226
+
227
+ https://developers.openai.com/api/docs/models/gpt-image-1.5
228
+
229
+ ## `"openai/gpt-image-1-mini": {`
230
+
231
+ https://developers.openai.com/api/docs/models/gpt-image-1-mini
232
+
233
+ ## `constraints: null,`
234
+
235
+ Preset-only — provider rejects arbitrary sizes.
236
+
237
+ ## `pricing: {`
238
+
239
+ https://developers.openai.com/api/docs/models/gpt-image-1-mini
240
+
241
+ ## `"google/gemini-3.1-flash-image-preview": {`
242
+
243
+ ---
244
+
245
+ Google (multimodal LLMs with native image output)
246
+ -----------------------------------------------------------------
247
+
248
+ python .tools/model_info.py --image gemini-3.1-flash-image
249
+ Vercel AI Gateway pricing: $0.50/MTok input, $3.00/MTok output
250
+
251
+ ## `providers: {`
252
+
253
+ "Nano Banana 2"; ids/prices verified 2026-06-29, see issues/908
254
+
255
+ ## `vercel: {`
256
+
257
+ Vercel AI Gateway serves the graduated `google/gemini-3.1-flash-image`
258
+ and the `-preview` alias at identical rates (feed, 2026-09-02).
259
+ Bindings call the graduated id; the canonical key above stays
260
+ `-preview` because it is persisted in selections and published
261
+ in the catalogue snapshot — renaming it is a separate change.
262
+
263
+ ## `references: { id: "google/gemini-3.1-flash-image", max: 14 },`
264
+
265
+ input_references advertised by OpenRouter (0–14), 2026-07-01.
266
+
267
+ ## `fal: {`
268
+
269
+ fal's graduated endpoint is `fal-ai/nano-banana-2`; same $0.08
270
+ per 1K image as the `-preview` endpoint (2K ×1.5, 4K ×2).
271
+
272
+ ## `"google/gemini-3-pro-image": {`
273
+
274
+ python .tools/model_info.py --image gemini-3-pro-image
275
+ Vercel AI Gateway pricing: $2.00/MTok input, $12.00/MTok output
276
+
277
+ ## `providers: {`
278
+
279
+ "Nano Banana Pro"; ids/prices verified 2026-06-29, see issues/908
280
+
281
+ ## `references: { id: "google/gemini-3-pro-image-preview", max: 14 },`
282
+
283
+ input_references advertised by OpenRouter (0–14), 2026-07-01.
284
+
285
+ ## `"google/gemini-3.1-flash-lite-image": {`
286
+
287
+ "Nano Banana 2 Lite" — GA 2026-06-30. The cost/speed tier of the 3.1
288
+ Flash family: ~half of Nano Banana 2's meter, and 1K-only output
289
+ (2K/4K unsupported — the differentiator). Vercel AI Gateway + OpenRouter both
290
+ meter it at $0.25/$1.50 (verified 2026-07-01); fal id not verified,
291
+ so left out. OpenRouter doesn't advertise input_references for the
292
+ Lite (t2i only per its model page), so no `references` (TOOL-DESIGN:
293
+ no unverified capability).
294
+
295
+ ## `avg_cost_usd: 0.034,`
296
+
297
+ Published 1K per-image cost (the card default): $0.034 = 1120 tokens
298
+ × $30/1M image-output (Google/Vercel AI Gateway changelog, 2026-07-01). The
299
+ budget meter charges this per image, so it must be the real cost.
300
+
301
+ ## `"bfl/flux-2-pro": {`
302
+
303
+ ---
304
+
305
+ Black Forest Labs (via Vercel AI Gateway)
306
+ -----------------------------------------------------------------
307
+
308
+ https://vercel.com/docs/ai-gateway/capabilities/image-generation/ai-sdk
309
+ https://docs.bfl.ml/pricing
310
+
311
+ ## `providers: {`
312
+
313
+ All three providers meter $0.03 per megapixel (Vercel AI Gateway model page,
314
+ OpenRouter endpoint `cost_usd`/megapixel, fal "first megapixel");
315
+ represented as flat at the 1MP baseline. The Vercel AI Gateway binding shipped
316
+ at $0.06 — a 2x hosted over-billing — corrected 2026-09-02. The
317
+ Vercel AI Gateway feed carries no `pricing` for BFL cards, so the page is the
318
+ source.
319
+
320
+ ## `references: { id: "black-forest-labs/flux.2-pro", max: 8 },`
321
+
322
+ input_references advertised by OpenRouter (0–8), 2026-07-01.
323
+
324
+ ## `"bfl/flux-2-max": {`
325
+
326
+ ---
327
+
328
+ Black Forest Labs — Flux 2 Max
329
+ -----------------------------------------------------------------
330
+
331
+ BFL's top Flux 2 line (2025-12-16). $0.07 per megapixel on all three
332
+ (Vercel AI Gateway model page — the feed carries no BFL pricing; OpenRouter
333
+ `cost_usd`/megapixel; fal "first megapixel", +$0.03 each additional).
334
+ Represented as flat at the 1MP baseline. Verified 2026-09-02.
335
+
336
+ ## `references: { id: "black-forest-labs/flux.2-max", max: 8 },`
337
+
338
+ OpenRouter `supported_parameters.input_references` 0–8
339
+ (2026-09-02). As on Flux 2 Pro.
340
+
341
+ ## `constraints: { min_edge: 256, max_edge: 1440 },`
342
+
343
+ Same envelope as Flux 2 Pro pending a vendor spec for Max.
344
+
345
+ ## `providers: {`
346
+
347
+ $0.04 on both providers: Vercel AI Gateway feed `pricing.image` + fal's model
348
+ page ("Fixed $0.04 cost per image edit"), verified 2026-09-02.
349
+
350
+ ## `"bytedance/seedream-5.0-pro": {`
351
+
352
+ ---
353
+
354
+ ByteDance — Seedream 5.0 Pro
355
+ -----------------------------------------------------------------
356
+
357
+ Vercel AI Gateway feed `pricing.image` $0.035 (the model page's rate table shows
358
+ $0.04 while its copy says $0.035 — the feed is the billing contract);
359
+ OpenRouter `cost_usd` $0.045 at 1K ($0.09 high-res, +$0.003 per
360
+ input image, not modelled); fal $0.0675 for ≤1536² area, $0.135 up
361
+ to 2048². Represented at the 1024² baseline. Verified 2026-09-02.
362
+
363
+ ## `references: { id: "bytedance-seed/seedream-5-0-pro", max: 14 },`
364
+
365
+ OpenRouter `supported_parameters.input_references` 0–14
366
+ (2026-09-02). Same endpoint as t2i, as on 4.5.
367
+
368
+ ## `constraints: {`
369
+
370
+ fal: total pixels between 1024x1024 and 2048x2048, aspect ratio
371
+ between 1:16 and 16:1 — an area bound, not a per-edge one, so a
372
+ 512x2048 request is in-envelope and a 2048x2048 one is the ceiling.
373
+
374
+ ## `"bytedance/seedream-5.0-lite": {`
375
+
376
+ ---
377
+
378
+ ByteDance — Seedream 5.0 Lite
379
+ -----------------------------------------------------------------
380
+
381
+ Universal: $0.035/img on all three (Vercel AI Gateway feed `pricing.image`,
382
+ OpenRouter `cost_usd`, fal page payload), verified 2026-09-02. A
383
+ 2K–4K model: fal scales requests below 2560x1440 up to its floor.
384
+
385
+ ## `references: { id: "bytedance-seed/seedream-5-0-lite", max: 14 },`
386
+
387
+ OpenRouter `supported_parameters.input_references` 0–14
388
+ (2026-09-02).
389
+
390
+ ## `constraints: { min_pixels: 3_686_400, max_pixels: 16_777_216 },`
391
+
392
+ fal: total pixels between 2560x1440 and 4096x4096 (requests below
393
+ the floor are scaled up to it); Vercel AI Gateway and OpenRouter serve 2K/4K
394
+ only. An area envelope, so the default is a 2K request.
395
+
396
+ ## `"bytedance/seedream-4.5": {`
397
+
398
+ ---
399
+
400
+ ByteDance — Seedream 4.5
401
+ -----------------------------------------------------------------
402
+
403
+ $0.04/img on all three providers (re-verified 2026-09-02).
404
+
405
+ ## `references: { id: "bytedance-seed/seedream-4.5", max: 14 },`
406
+
407
+ i2i verified live 2026-07-01 (OpenRouter /api/v1/images
408
+ input_references; same endpoint as t2i). max from
409
+ supported_parameters.input_references.
410
+
411
+ ## `"xai/grok-imagine-image-2.0": {`
412
+
413
+ ---
414
+
415
+ SpaceXAI — Grok Imagine Image 2.0
416
+ -----------------------------------------------------------------
417
+
418
+ Tiered by quality (low/medium) × resolution (1K/2K); the same four
419
+ rates on all three providers (Vercel AI Gateway feed
420
+ `image_dimension_quality_pricing`, OpenRouter `cost_usd` variants,
421
+ fal page). OpenRouter also bills $0.01 per input image (not
422
+ modelled). Verified 2026-09-02.
423
+
424
+ ## `references: { id: "x-ai/grok-imagine-image-2.0", max: 3 },`
425
+
426
+ OpenRouter `supported_parameters.input_references` 0–3
427
+ (2026-09-02). OR bills $0.01 per input image on top (not modelled).
428
+
429
+ ## `"meta/muse-image-1.0": {`
430
+
431
+ ---
432
+
433
+ Meta — Muse Image 1.0
434
+ -----------------------------------------------------------------
435
+
436
+ Meta's agentic image model (2026-08-26): $0.01/img on Vercel AI Gateway (feed
437
+ `pricing.image`) and fal (page payload). OpenRouter lists it but
438
+ exposes no serving endpoint.
439
+ fal exposes aspect ratio only (no size control). Verified 2026-09-02.
440
+
441
+ ## `constraints: null,`
442
+
443
+ Muse picks output dimensions from the aspect ratio; neither
444
+ provider exposes a size envelope, so none is claimed.
445
+
446
+ ## `"recraft/recraft-v4.1": {`
447
+
448
+ ---
449
+
450
+ Recraft — V4.1
451
+ -----------------------------------------------------------------
452
+
453
+ Universal: $0.035/img raster on every provider (Vercel AI Gateway feed
454
+ `pricing.image`, OpenRouter endpoint `cost_usd`, fal page payload),
455
+ verified 2026-09-02. Vector styles are $0.08 and a separate route on
456
+ fal/OpenRouter (`.../text-to-vector`, `recraft-v4.1-vector`) — not
457
+ catalogued; `styles: null` here means the raster route only.
458
+ OpenRouter org slug is `recraft`, not `recraft-ai`.
459
+
460
+ ## `references: { id: "recraft/recraft-v4.1", max: 1 },`
461
+
462
+ OpenRouter `supported_parameters.input_references` 0–1
463
+ (2026-09-02).
464
+
465
+ ## `"recraft/recraft-v3": {`
466
+
467
+ ---
468
+
469
+ Recraft — V3
470
+ -----------------------------------------------------------------
471
+
472
+ $0.04/img raster on every provider (re-verified 2026-09-02; Recraft's
473
+ own pricing table agrees).
474
+
475
+ ## `pricing: { type: "per_run_flat", usd: 0.04 },`
476
+
477
+ Source: replicate.com/google/lyria-3 — "$0.04 per output audio file"
478
+
479
+ ## `pricing: { type: "per_run_flat", usd: 0.08 },`
480
+
481
+ Source: replicate.com/google/lyria-3-pro — "$0.08 per output audio file"
482
+
483
+ ## `pricing: {`
484
+
485
+ API: 100 credits when duration is automatic, or 11 credits/s when set.
486
+
487
+ ## `pricing: { type: "per_1000_characters", usd: 0.1 },`
488
+
489
+ Source: elevenlabs.io/pricing/api — $0.10 per 1,000 characters.
490
+
491
+ ## `"tripo/rig-v2.5": {`
492
+
493
+ Follow the documented compatibility table, not its contradictory biped example.
494
+
495
+ ## `"google/veo-3.1": {`
496
+
497
+ ---
498
+
499
+ Google — Veo 3.1
500
+ -----------------------------------------------------------------
501
+
502
+ ## `vercel: {`
503
+
504
+ Vercel AI Gateway — vercelAiGateway.videoModel(id), image-to-video. The
505
+ Vercel AI Gateway meters both audio modes and sells 4K; the matrix is
506
+ identical to fal's. Verified against Vercel AI Gateway's own
507
+ /v1/models feed (`video_duration_pricing`) on 2026-09-02 — the
508
+ previous card claimed "audio-on only, ≤1080p", which the feed
509
+ contradicts.
510
+
511
+ `silent` is catalogued as provider truth but is not yet
512
+ reachable on Grida's hosted route: the video wire carries no
513
+ audio-mode field, so `generateVideo` prices the card's default
514
+ mode ([`editor/lib/ai/server.ts`](https://github.com/gridaco/grida/blob/main/editor/lib/ai/server.ts) in gridaco/grida). `"4k"` IS reachable — the
515
+ request path maps a 2160 short edge to that label and today
516
+ rejects it for want of a rate.
517
+ https://vercel.com/ai-gateway/models/veo-3.1-generate-001
518
+
519
+ ## `fal: {`
520
+
521
+ fal.ai — image-to-video endpoint (capability is keyed into the id;
522
+ t2v is a separate `fal-ai/veo3.1` endpoint, not catalogued). Rate
523
+ matrix is identical to the Vercel AI Gateway's: $0.40/s audio and
524
+ $0.20/s silent at 720p/1080p, $0.60/$0.40 at 4K.
525
+ https://fal.ai/models/fal-ai/veo3.1/image-to-video
526
+
527
+ ## `openrouter: {`
528
+
529
+ OpenRouter — async `/api/v1/videos` (job → poll → unsigned url).
530
+ Text-to-video + image-to-video; native audio. "from $0.40/s"
531
+ (verified 2026-06-29, https://openrouter.ai/google/veo-3.1).
532
+
533
+ ## `"google/veo-3.1-fast": {`
534
+
535
+ ---
536
+
537
+ Google — Veo 3.1 Fast
538
+ -----------------------------------------------------------------
539
+
540
+ Same envelope as Veo 3.1 (16:9/9:16, 4/6/8s, native audio, ≤4K) at
541
+ ~2.7x lower cost. Rate matrix is identical on Vercel AI Gateway and fal —
542
+ Vercel AI Gateway feed `video_duration_pricing` and fal's stated per-second
543
+ rates, verified 2026-09-02.
544
+
545
+ ## `"google/veo-3.1-lite": {`
546
+
547
+ ---
548
+
549
+ Google — Veo 3.1 Lite
550
+ -----------------------------------------------------------------
551
+
552
+ The budget Veo: 720p/1080p only, 4/6/8s, native audio. Rates
553
+ identical on Vercel AI Gateway and fal (verified 2026-09-02). Vercel AI Gateway
554
+ supports both text and image input (FAQ verified 2026-09-07); the fal binding
555
+ below requires an image. Hosted GG's narrower wire is a host concern.
556
+
557
+ ## `"alibaba/wan-3.0": {`
558
+
559
+ ---
560
+
561
+ Alibaba — Wan 3.0
562
+ -----------------------------------------------------------------
563
+
564
+ Per-second by resolution, audio bundled into the rate (Vercel AI Gateway
565
+ feed has no audio axis; fal exposes an `audio` toggle but bills the
566
+ same). Identical on Vercel AI Gateway and fal, verified 2026-09-02. 2–30s.
567
+
568
+ ## `"bytedance/seedance-2.0": {`
569
+
570
+ ByteDance — Seedance 2.0
571
+ -----------------------------------------------------------------
572
+
573
+ ## `providers: {`
574
+
575
+ NO Vercel AI Gateway binding, deliberately. Vercel AI Gateway serves
576
+ `bytedance/seedance-2.0` but meters it PER TOKEN
577
+ (`video_token_pricing`: $7.00/MTok at 480p/720p, $7.70 at 1080p,
578
+ $4.00 at 4K; reduced with video input; "minimum token floors based
579
+ on output duration"). `PerSecondPricing` cannot express that, the
580
+ hosted route pre-prices rate×duration, and ByteDance publishes no
581
+ tokens-per-second figure — so there is no honest per-second rate,
582
+ and the per-second Vercel AI Gateway binding this card shipped with was
583
+ invented. Withheld until the hosted path can meter post-flight;
584
+ hosted requests fail with "not available on the hosted provider".
585
+ See gridaco/grida#1019 (Blocker A). Feed checked 2026-09-02.
586
+
587
+ ## `fal: {`
588
+
589
+ fal — image-to-video. fal also meters tokens underneath
590
+ ($0.014/1K), but states a per-second price for the two
591
+ resolutions it prices: $0.3034/s @720p, $0.682/s @1080p, audio
592
+ bundled. Those are fal's own figures, not a conversion of ours.
593
+ https://fal.ai/models/bytedance/seedance-2.0/image-to-video
594
+
595
+ ## `openrouter: {`
596
+
597
+ OpenRouter — async `/api/v1/videos`. Flat $0.06726/s (verified
598
+ 2026-06-29, https://openrouter.ai/bytedance/seedance-2.0) — far
599
+ below Vercel AI Gateway's per-resolution rate (the proprietary-pricing-
600
+ diverges finding, #908). No separate silent meter surfaced.
601
+
602
+ ## `"bytedance/seedance-2.5": {`
603
+
604
+ ---
605
+
606
+ ByteDance — Seedance 2.5
607
+ -----------------------------------------------------------------
608
+
609
+ Newer generation (2026-07-31), NOT a drop-in successor: ~55% more
610
+ per token on Vercel AI Gateway ($10.70/MTok at 480p/720p, $11.70 at 1080p
611
+ vs 2.0's $7.00/$7.70), ~56–71% more per second on fal, and no 4K.
612
+ It buys 4–30s clips, video-editing and extend-video. 2.0 is cheaper
613
+ and serves 4K. Vercel AI Gateway bills tokens, so there is no comparable
614
+ per-second Vercel AI Gateway meter here. fal states per-second prices, audio bundled.
615
+
616
+ ## `fal: {`
617
+
618
+ fal's stated rates: $0.2205/s @480p, $0.4730/s @720p, $1.164/s
619
+ @1080p (fal meters $0.0214/1K tokens underneath). 2026-09-02.
620
+ https://fal.ai/models/bytedance/seedance-2.5/image-to-video
621
+
622
+ ## `"xai/grok-imagine-video-1.5": {`
623
+
624
+ ---
625
+
626
+ SpaceXAI — Grok Imagine Video 1.5
627
+ -----------------------------------------------------------------
628
+
629
+ Image-to-video only (no t2v, per SpaceXAI docs); native lip-synced audio
630
+ bundled into the rate. Per-second by resolution, identical on Vercel AI Gateway
631
+ (no markup) and fal: $0.08/s @480p, $0.14/s @720p, $0.25/s @1080p.
632
+ Both also bill $0.01 per input image, captured separately from the
633
+ output meter.
634
+
635
+ ## `vercel: {`
636
+
637
+ Vercel AI Gateway — image-to-video; mirrors SpaceXAI's list price (no markup).
638
+ Vercel AI Gateway namespaces every SpaceXAI model under `spacexai/`, so the
639
+ call id deliberately differs from this card's canonical `xai/` id.
640
+ https://vercel.com/changelog/grok-imagine-video-1-5-on-ai-gateway
641
+
642
+ ## `fal: {`
643
+
644
+ fal.ai — image-to-video endpoint; same per-second rate.
645
+ https://fal.ai/models/xai/grok-imagine-video/v1.5/image-to-video
646
+
647
+ ## `pricing: { type: "per_token_input", input: 0.15 },`
648
+
649
+ VERIFY before prod: models.dev lists no per-1M price for
650
+ gemini-embedding-2 yet; using the gemini-embedding-001 family
651
+ input rate ($0.15 / 1M) as a conservative stand-in.