@grida/ai-models 1.20261011.4 → 1.20261013.1
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/PROVENANCE.md +311 -13
- package/README.md +2 -2
- package/manifest.json +13 -13
- package/models.json +12604 -198
- package/package.json +1 -1
- package/schemas/models.schema.json +891 -13
package/PROVENANCE.md
CHANGED
|
@@ -101,9 +101,8 @@ https://platform.claude.com/docs/en/models/sonnet-5-5/whats-new-sonnet-5-5
|
|
|
101
101
|
|
|
102
102
|
## `"anthropic/claude-sonnet-5": {`
|
|
103
103
|
|
|
104
|
-
$2/$10
|
|
105
|
-
|
|
106
|
-
announced step up to $3/$15. Do not restore the higher card.
|
|
104
|
+
$2/$10 per million input/output tokens is Anthropic's published regular rate.
|
|
105
|
+
Do not substitute an obsolete higher rate.
|
|
107
106
|
https://platform.claude.com/docs/en/about-claude/pricing
|
|
108
107
|
|
|
109
108
|
## `"anthropic/claude-fable-5.1": {`
|
|
@@ -134,11 +133,12 @@ https://platform.claude.com/docs/en/models/opus-5-5/migration-guide
|
|
|
134
133
|
Google's cache model is read + hourly storage (no one-time write
|
|
135
134
|
premium that matches `cacheWrite` semantics), so the field is omitted.
|
|
136
135
|
|
|
137
|
-
|
|
138
|
-
|
|
139
|
-
|
|
140
|
-
|
|
141
|
-
and this exact id is
|
|
136
|
+
The intrinsic cost records Google's published regular rates: $1.50 input,
|
|
137
|
+
$7.50 output, and $0.15 cached input per million tokens. Promotional rates
|
|
138
|
+
and their expiry dates are excluded. This evidence applies to Google's
|
|
139
|
+
rates; each serving provider's regular prices need its own evidence.
|
|
140
|
+
Gemini 3.8 Flash is GA, and this exact id is listed on both Vercel AI Gateway
|
|
141
|
+
and OpenRouter.
|
|
142
142
|
https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash
|
|
143
143
|
https://ai.google.dev/gemini-api/docs/pricing
|
|
144
144
|
https://vercel.com/ai-gateway/models/gemini-3.8-flash
|
|
@@ -391,9 +391,11 @@ Represented as flat at the 1MP baseline. Verified 2026-09-02.
|
|
|
391
391
|
OpenRouter `supported_parameters.input_references` 0–8
|
|
392
392
|
(2026-09-02). As on Flux 2 Pro.
|
|
393
393
|
|
|
394
|
-
## `constraints:
|
|
394
|
+
## `constraints: null,`
|
|
395
395
|
|
|
396
|
-
|
|
396
|
+
The inherited Flux 2 Pro edge envelope was unverified for Max and is withdrawn.
|
|
397
|
+
See the 2026-10-13 media qualifications below for the independently sourced
|
|
398
|
+
fal endpoint bounds.
|
|
397
399
|
|
|
398
400
|
## `providers: {`
|
|
399
401
|
|
|
@@ -571,8 +573,8 @@ https://vercel.com/ai-gateway/models/veo-3.1-generate-001
|
|
|
571
573
|
|
|
572
574
|
## `fal: {`
|
|
573
575
|
|
|
574
|
-
fal.ai — image-to-video endpoint
|
|
575
|
-
|
|
576
|
+
fal.ai — image-to-video endpoint `fal-ai/veo3.1/image-to-video`; the
|
|
577
|
+
separate text-to-video endpoint `fal-ai/veo3.1` is also catalogued. Rate
|
|
576
578
|
matrix is identical to the Vercel AI Gateway's: $0.40/s audio and
|
|
577
579
|
$0.20/s silent at 720p/1080p, $0.60/$0.40 at 4K.
|
|
578
580
|
https://fal.ai/models/fal-ai/veo3.1/image-to-video
|
|
@@ -605,7 +607,7 @@ Google — Veo 3.1 Lite
|
|
|
605
607
|
The budget Veo: 720p/1080p only, 4/6/8s, native audio. Rates
|
|
606
608
|
identical on Vercel AI Gateway and fal (verified 2026-09-02). Vercel AI Gateway
|
|
607
609
|
supports both text and image input (FAQ verified 2026-09-07); the fal binding
|
|
608
|
-
|
|
610
|
+
requires an image.
|
|
609
611
|
|
|
610
612
|
## `"alibaba/wan-3.0": {`
|
|
611
613
|
|
|
@@ -702,3 +704,299 @@ https://fal.ai/models/xai/grok-imagine-video/v1.5/image-to-video
|
|
|
702
704
|
VERIFY before prod: models.dev lists no per-1M price for
|
|
703
705
|
gemini-embedding-2 yet; using the gemini-embedding-001 family
|
|
704
706
|
input rate ($0.15 / 1M) as a conservative stand-in.
|
|
707
|
+
|
|
708
|
+
## Text labs and Vercel endpoints (2026-10-13)
|
|
709
|
+
|
|
710
|
+
Text cards identify the lab named by their existing first-party release source:
|
|
711
|
+
OpenAI, Anthropic, or Google. Their exact serving IDs are verified against
|
|
712
|
+
[Vercel's public model discovery](https://ai-gateway.vercel.sh/v1/models),
|
|
713
|
+
which lists all 17 recorded text models. These identifiers record availability,
|
|
714
|
+
not a routing preference. Existing model release dates are unchanged. Standard
|
|
715
|
+
serving prices, bands and limits are qualified separately below.
|
|
716
|
+
|
|
717
|
+
## fal input variants (2026-10-13)
|
|
718
|
+
|
|
719
|
+
Bindings list the text and image variants documented on their exact fal API
|
|
720
|
+
pages. Each endpoint includes its own source URL; existing `references` fields
|
|
721
|
+
remain unchanged. GPT Image 2's ID is corrected to `openai/gpt-image-2` from
|
|
722
|
+
[fal's API example](https://fal.ai/models/openai/gpt-image-2/api); the overview's
|
|
723
|
+
hand-written quick start still shows `fal-ai/gpt-image-2`, which disagrees with
|
|
724
|
+
the generated API documentation. The generated endpoint-specific API example
|
|
725
|
+
is the more specific source. This correction does not change prices.
|
|
726
|
+
|
|
727
|
+
## Tripo rig check
|
|
728
|
+
|
|
729
|
+
[Tripo's rig-check documentation](https://developers.tripo3d.ai/en/docs/animations-rig-check)
|
|
730
|
+
identifies the operation and lab but does not establish its first public release
|
|
731
|
+
day. It carries `basis: provider_endpoint` and `date: null`, as do the rigging
|
|
732
|
+
endpoints; a consumer must not substitute its own integration date.
|
|
733
|
+
|
|
734
|
+
## Media endpoint parameters and limits (2026-10-13)
|
|
735
|
+
|
|
736
|
+
The recorded fal image, video and 3D variants now include native parameter facts
|
|
737
|
+
from their endpoint-specific API pages and public OpenAPI documents. Each
|
|
738
|
+
endpoint's `source_url` identifies that variant; a parameter with a different
|
|
739
|
+
source carries its own `source_url`. Names such as `image_size.width` denote
|
|
740
|
+
native nested fields. An `image_size` preset enum is the published string-option
|
|
741
|
+
list, not a complete validator for its string-or-object input.
|
|
742
|
+
|
|
743
|
+
Model-specific `x-fal` size metadata takes precedence over generic `ImageSize`
|
|
744
|
+
component bounds. The generic component alone does not establish a model's
|
|
745
|
+
intrinsic envelope. Area and aspect bounds are recorded only where the exact
|
|
746
|
+
endpoint publishes them. In particular:
|
|
747
|
+
|
|
748
|
+
- [Seedream 5.0 Lite's input schema](https://fal.ai/models/bytedance/seedream/v5/lite/text-to-image/api)
|
|
749
|
+
explicitly states a 2560×1440 to 4096×4096 total-area range. Its below-floor
|
|
750
|
+
scaling behavior is separate from that supported output range.
|
|
751
|
+
- Seedream edit's ten references and Grok Image edit's five references come from
|
|
752
|
+
their exact endpoint schema descriptions, which are more specific than the
|
|
753
|
+
missing JSON `maxItems` keywords. Nano Banana 2 and Pro's fourteen-reference
|
|
754
|
+
limits come from their respective [Nano Banana 2 edit overview](https://fal.ai/models/fal-ai/nano-banana-2/edit)
|
|
755
|
+
and [Nano Banana Pro edit overview](https://fal.ai/models/fal-ai/nano-banana-pro/edit).
|
|
756
|
+
- [Flux Pro's edit overview](https://fal.ai/models/fal-ai/flux-2-pro/edit) publishes
|
|
757
|
+
nine references and 9MP total input. Flux Max's corresponding reference count
|
|
758
|
+
is unspecified. Neither edit schema publishes the generation variant's
|
|
759
|
+
`x-fal` dimension extension, so those generation bounds are not copied to edit.
|
|
760
|
+
- Flux and Recraft do not publish a `num_images` parameter in these endpoint
|
|
761
|
+
schemas. Unpublished seed bounds and reference counts remain absent. A missing
|
|
762
|
+
fact does not establish unlimited input or an intrinsic model limit.
|
|
763
|
+
|
|
764
|
+
### Intrinsic envelope corrections
|
|
765
|
+
|
|
766
|
+
[BytePlus's Seedance 2.0 documentation](https://docs.byteplus.com/en/docs/modelark/seedance-2-0)
|
|
767
|
+
and its [video-generation tutorial](https://docs.byteplus.com/id/docs/modelark/video-generation-tutorial)
|
|
768
|
+
establish the 4–15-second duration range and six aspect ratios: 21:9, 16:9, 4:3,
|
|
769
|
+
1:1, 3:4 and 9:16. The old five-second minimum and incomplete ratio list are
|
|
770
|
+
corrected. The exact fal text and image variants independently agree.
|
|
771
|
+
|
|
772
|
+
Gemini 3 Pro Image's inherited 1536-pixel edge limit is withdrawn. Its
|
|
773
|
+
[fal generation schema](https://fal.ai/models/fal-ai/nano-banana-pro/api) and
|
|
774
|
+
[edit schema](https://fal.ai/models/fal-ai/nano-banana-pro/edit/api) document
|
|
775
|
+
1K/2K/4K resolution choices. Those tier labels do not establish a universal
|
|
776
|
+
per-edge maximum: dimensions also depend on aspect ratio. No replacement
|
|
777
|
+
intrinsic envelope is inferred from another model's output-size table.
|
|
778
|
+
|
|
779
|
+
Flux Max's copied Pro edge envelope is also withdrawn. The
|
|
780
|
+
[lab overview](https://docs.bfl.ml/flux_2/flux2_overview) says up to 4MP without
|
|
781
|
+
establishing an exact pixel conversion. The
|
|
782
|
+
[fal generation schema](https://fal.ai/models/fal-ai/flux-2-max/api) independently
|
|
783
|
+
publishes 256–2560 per edge, steps of 16, and 4,194,304 total pixels. These are
|
|
784
|
+
provider-endpoint facts, not a conversion of the lab's 4MP statement.
|
|
785
|
+
|
|
786
|
+
### Variable fal image prices
|
|
787
|
+
|
|
788
|
+
The existing baseline `usd`, `type` and average-cost fields are unchanged.
|
|
789
|
+
Additive fields describe the published variable portions of those schedules:
|
|
790
|
+
|
|
791
|
+
- [Flux Pro](https://fal.ai/models/fal-ai/flux-2-pro/edit) bills $0.03 for the first
|
|
792
|
+
output megapixel and $0.015 per additional input/output megapixel, rounded up.
|
|
793
|
+
Its examples price 1024×1024 at $0.03 and 1920×1080 at $0.045. The first-unit
|
|
794
|
+
scope is `output`; no precise universal pixels-per-megapixel conversion is
|
|
795
|
+
inferred from those examples.
|
|
796
|
+
- [Flux Max](https://fal.ai/models/fal-ai/flux-2-max/edit) bills $0.07 for the first
|
|
797
|
+
processed megapixel and $0.03 for each additional megapixel, including inputs.
|
|
798
|
+
Its first-unit scope is `processed`. No rounding rule is copied from Pro.
|
|
799
|
+
- [Nano Banana 2](https://fal.ai/models/fal-ai/nano-banana-2) retains $0.08 per image.
|
|
800
|
+
Published multipliers are 0.75× for 0.5K, 1.5× for 2K and 2× for 4K. Web search
|
|
801
|
+
adds $0.015 and high thinking adds $0.002 when used. The
|
|
802
|
+
[edit page](https://fal.ai/models/fal-ai/nano-banana-2/edit) states the same terms.
|
|
803
|
+
- [Nano Banana Pro](https://fal.ai/models/fal-ai/nano-banana-pro) retains $0.15 per
|
|
804
|
+
image; 4K costs 2× and web search adds $0.015 when used. Its
|
|
805
|
+
[edit page](https://fal.ai/models/fal-ai/nano-banana-pro/edit) agrees. No additional
|
|
806
|
+
thinking fee or special 2K multiplier is inferred.
|
|
807
|
+
|
|
808
|
+
### Native 3D operation facts
|
|
809
|
+
|
|
810
|
+
Tripo generation, rigging and rig-check variants carry their own native API paths,
|
|
811
|
+
version enums and parameter facts. The sources are Tripo's separate
|
|
812
|
+
[P text-generation](https://developers.tripo3d.ai/en/docs/generation-text-to-model/p),
|
|
813
|
+
[P image-generation](https://developers.tripo3d.ai/en/docs/generation-image-to-model/p),
|
|
814
|
+
[P multiview](https://developers.tripo3d.ai/en/docs/generation-multiview-to-model/p),
|
|
815
|
+
[standard text-generation](https://developers.tripo3d.ai/en/docs/generation-text-to-model/standard),
|
|
816
|
+
[standard image-generation](https://developers.tripo3d.ai/en/docs/generation-image-to-model/standard),
|
|
817
|
+
[standard multiview](https://developers.tripo3d.ai/en/docs/generation-multiview-to-model/standard),
|
|
818
|
+
[rigging](https://developers.tripo3d.ai/en/docs/animations-rig) and
|
|
819
|
+
[rig-check](https://developers.tripo3d.ai/en/docs/animations-rig-check) pages.
|
|
820
|
+
Multiview is image input; its native path distinguishes the operation. Its
|
|
821
|
+
task-ID, four-slot legacy and named-view input forms do not justify one universal
|
|
822
|
+
array cardinality. H3.1's 2,000,000-face detailed bound does not replace its
|
|
823
|
+
1,500,000-face standard bound. No new seed limits or undocumented output-format
|
|
824
|
+
parameter are invented. The `fast` texture option requires the documented
|
|
825
|
+
`v3.5-20260815` texture version.
|
|
826
|
+
|
|
827
|
+
The fal [Hunyuan image](https://fal.ai/models/fal-ai/hunyuan-3d/v3.1/pro/image-to-3d/api),
|
|
828
|
+
[Hunyuan text](https://fal.ai/models/fal-ai/hunyuan-3d/v3.1/pro/text-to-3d/api) and
|
|
829
|
+
[TRELLIS.2](https://fal.ai/models/fal-ai/trellis-2/api) endpoints likewise record
|
|
830
|
+
their native input types, enums and bounds. Hunyuan's image-dimension prose
|
|
831
|
+
describes the reference input; it is not an output-image envelope. TRELLIS.2's
|
|
832
|
+
resolution, mesh, texture, sampling, guidance and UV parameters retain their
|
|
833
|
+
published units and bounds. These partial parameter maps do not claim to replace
|
|
834
|
+
the providers' complete request schemas or cross-field validation rules.
|
|
835
|
+
|
|
836
|
+
## Published image output calculators and examples (2026-10-13)
|
|
837
|
+
|
|
838
|
+
The GPT Image 2 and GPT Image 2.5 cards record the calculator in
|
|
839
|
+
[OpenAI's image generation guide](https://developers.openai.com/api/docs/guides/image-generation),
|
|
840
|
+
including its [published JavaScript](https://developers.openai.com/_astro/GptImageTokenCalculator.react.DukR73TZ.js).
|
|
841
|
+
That source supplies the quality coefficients, area constants, ties-to-even
|
|
842
|
+
short-axis rounding and final upward token rounding. The examples are estimates
|
|
843
|
+
from that calculator, not measured usage or guaranteed whole-request prices.
|
|
844
|
+
They exclude text/image inputs, reasoning and partial-image output charges.
|
|
845
|
+
Exact output-token examples were verified with the calculator rather than inferred
|
|
846
|
+
by dividing a rounded USD price by the token rate.
|
|
847
|
+
|
|
848
|
+
The GPT Image 2.5 cards also retain all six published size rows from fal's
|
|
849
|
+
[Flare](https://fal.ai/models/openai/gpt-image-2.5/flare/text-to-image) and
|
|
850
|
+
[Sunburst](https://fal.ai/models/openai/gpt-image-2.5/sunburst/text-to-image)
|
|
851
|
+
price tables. Their `1920x1080` row conflicts with the documented requirement that
|
|
852
|
+
both dimensions be divisible by 16: 1080 is not. That quoted USD example remains
|
|
853
|
+
source-backed, but has no `output_tokens` value. OpenAI's calculator UI rejects
|
|
854
|
+
that size; calling only its unvalidated arithmetic function is not evidence of
|
|
855
|
+
an admissible request. Example dimensions must therefore be checked against the
|
|
856
|
+
model and endpoint constraints before use.
|
|
857
|
+
|
|
858
|
+
[fal's GPT Image 2 page](https://fal.ai/models/openai/gpt-image-2) bills text and
|
|
859
|
+
image tokens, so its binding now records token rates instead of treating the
|
|
860
|
+
rounded per-image examples as a fixed tariff. All three fal GPT Image bindings
|
|
861
|
+
publish the same six native token rates and say total cost rounds upward to
|
|
862
|
+
`$0.0001`, recorded as `rounding_usd`. A rounded example is not an exact total
|
|
863
|
+
request maximum. OpenAI's separately published cached-input rates apply to its
|
|
864
|
+
Responses image-generation tool, not direct Images API requests; recording a
|
|
865
|
+
provider's cache price does not promise a cache hit or cached billing on every API.
|
|
866
|
+
|
|
867
|
+
The six fal image endpoints' option bounds come from their endpoint-specific
|
|
868
|
+
public OpenAPI schemas. Image 2 allows at most four outputs; Image 2.5 allows ten.
|
|
869
|
+
Prompt bounds are character counts, not bytes or tokens. The shared generic
|
|
870
|
+
`ImageSize` definition permits larger dimensions, but the model-specific size
|
|
871
|
+
description and OpenAI guide impose the stricter 3840px edge, pixel-area,
|
|
872
|
+
16px-step and aspect constraints recorded on the endpoints. No undocumented
|
|
873
|
+
minimum number of reference images is inferred from an array being required.
|
|
874
|
+
|
|
875
|
+
## Vision tokenization parameters (2026-10-13)
|
|
876
|
+
|
|
877
|
+
[OpenAI's vision guide](https://developers.openai.com/api/docs/guides/images-vision)
|
|
878
|
+
explicitly lists GPT-5.5, GPT-5.6 Sol/Terra/Luna and GPT-6 Astra as 32px patch
|
|
879
|
+
models with a 1.2 token multiplier. It also supplies detail-specific resizing
|
|
880
|
+
limits and the separate 30,000-patch rejection threshold. Low detail fits within
|
|
881
|
+
512px on each side; a 256-patch allowance can be derived from that square, but is
|
|
882
|
+
not recorded as an independently published patch budget. The guide does not
|
|
883
|
+
qualify GPT-5.5 Pro's patch formula, so that card has no invented formula.
|
|
884
|
+
|
|
885
|
+
[Claude's vision guide](https://platform.claude.com/docs/en/build-with-claude/vision)
|
|
886
|
+
identifies 28px visual-token patches, a 2576px long edge and a 4784 visual-token
|
|
887
|
+
ceiling for Claude 4.7 and later. These are processing limits: oversized native
|
|
888
|
+
images may be downscaled before tokenization.
|
|
889
|
+
|
|
890
|
+
[Google's media-resolution guide](https://ai.google.dev/gemini-api/docs/media-resolution)
|
|
891
|
+
labels the Gemini 3 family token counts approximate. The cards preserve that
|
|
892
|
+
qualification and each resolution's published budget, including the 2240-token
|
|
893
|
+
`ultra_high` budget. That maximum is not a fixed billable count for every image.
|
|
894
|
+
These parameters describe provider tokenization. Preflight estimates do not
|
|
895
|
+
establish final charges.
|
|
896
|
+
|
|
897
|
+
## Text provider references and regular pricing (2026-10-13)
|
|
898
|
+
|
|
899
|
+
All 17 text cards retain independent native, Vercel and OpenRouter bindings.
|
|
900
|
+
The native model pages establish exact IDs and synchronous limits. Claude's
|
|
901
|
+
[pricing table](https://platform.claude.com/docs/en/about-claude/pricing) separates
|
|
902
|
+
five-minute and one-hour cache writes. Google's
|
|
903
|
+
[Standard pricing](https://ai.google.dev/gemini-api/docs/pricing) excludes optional
|
|
904
|
+
cache storage, grounding and alternate service tiers. Native and intermediary
|
|
905
|
+
limits and input-length bands intentionally remain distinct; a native reference
|
|
906
|
+
is a presentation choice, not a routing preference.
|
|
907
|
+
|
|
908
|
+
[Vercel's model feed](https://ai-gateway.vercel.sh/v1/models) supplies serving
|
|
909
|
+
limits and supported capabilities. Prices were checked against each model's
|
|
910
|
+
public page for regular-rate qualification. The
|
|
911
|
+
[Gemini 3.7 Flash](https://vercel.com/ai-gateway/models/gemini-3.7-flash) and
|
|
912
|
+
[Gemini 3.8 Flash](https://vercel.com/ai-gateway/models/gemini-3.8-flash) pages
|
|
913
|
+
publish explicit list tiers of $1.50 input, $7.50 output and $0.15 cached input
|
|
914
|
+
per million tokens. These replace discounted feed values; they are not inferred
|
|
915
|
+
by reversing a percentage. OpenRouter's corresponding model pages independently
|
|
916
|
+
publish the same list rates. Standard/global pricing excludes regional premiums,
|
|
917
|
+
optional tools and alternate service tiers.
|
|
918
|
+
|
|
919
|
+
[OpenRouter's model feed](https://openrouter.ai/api/v1/models) supplies its exact
|
|
920
|
+
IDs, modalities, limits and capabilities. Each model page qualifies the rates;
|
|
921
|
+
provider input-length boundaries are preserved exactly. The
|
|
922
|
+
[GPT-5.6 Sol endpoint listing](https://openrouter.ai/api/v1/models/openai/gpt-5.6-sol/endpoints)
|
|
923
|
+
independently establishes the recorded Standard rates and tiers. Time-based
|
|
924
|
+
cache-storage rates are not per-token cache-write prices.
|
|
925
|
+
|
|
926
|
+
The native [GPT-5.6 Sol page](https://developers.openai.com/api/docs/models/gpt-5.6-sol)
|
|
927
|
+
and [OpenAI pricing](https://developers.openai.com/api/docs/pricing) qualify the
|
|
928
|
+
visible native rate as promotional without publishing a regular replacement.
|
|
929
|
+
That binding therefore omits pricing; unknown does not mean free. Its Vercel and
|
|
930
|
+
OpenRouter prices remain independently sourced. No discount schedule is tracked.
|
|
931
|
+
GPT-5.5 Pro has no invented native cache discount or unverified native
|
|
932
|
+
long-context price; intermediary bands remain separate provider facts.
|
|
933
|
+
|
|
934
|
+
## TRELLIS.2 distribution (2026-10-13)
|
|
935
|
+
|
|
936
|
+
Microsoft's [official repository](https://github.com/microsoft/TRELLIS.2) links
|
|
937
|
+
the [TRELLIS.2-4B weights](https://huggingface.co/microsoft/TRELLIS.2-4B).
|
|
938
|
+
The [code license](https://github.com/microsoft/TRELLIS.2/blob/main/LICENSE)
|
|
939
|
+
and the weight model card both identify MIT; these are recorded separately.
|
|
940
|
+
The Hugging Face model metadata reports `gated: false`, so the published weight
|
|
941
|
+
access is recorded as public. Dependency licenses remain separate from these
|
|
942
|
+
artifact licenses. Distribution links do not establish a hosted inference API,
|
|
943
|
+
free inference or permission for every possible use.
|
|
944
|
+
|
|
945
|
+
## Provider coverage and non-text differences (2026-10-13)
|
|
946
|
+
|
|
947
|
+
All existing cards are retained. Provider coverage was checked against
|
|
948
|
+
[OpenRouter discovery](https://openrouter.ai/api/v1/models) and its image, video
|
|
949
|
+
and embedding feeds, [Vercel discovery](https://ai-gateway.vercel.sh/v1/models),
|
|
950
|
+
and fal's endpoint catalogue. Exact aliases matter: OpenRouter's Muse and Lyria
|
|
951
|
+
preview IDs differ from canonical card IDs. `not_listed` records absence from
|
|
952
|
+
those inspected public listings, not a promise that a provider will never offer
|
|
953
|
+
an equivalent endpoint. fal coverage is restricted to media.
|
|
954
|
+
|
|
955
|
+
Native references now identify actual lab APIs, including BytePlus for ByteDance,
|
|
956
|
+
Google Developer API, Meta, BFL, Recraft, xAI and Alibaba. Open weights do not
|
|
957
|
+
invent a lab-hosted API. Native and reseller prices retain their own meters:
|
|
958
|
+
|
|
959
|
+
- OpenRouter GPT Image 2 uses separate text-input, image-input and image-output
|
|
960
|
+
token meters. Its 2.5 endpoints now list transparency; its Image 2 endpoint
|
|
961
|
+
lists only `auto` and `opaque`.
|
|
962
|
+
- Seedance video-token SKUs replace an inaccurate OpenRouter per-second estimate.
|
|
963
|
+
Input-video, resolution and audio variants remain distinct. BytePlus's current
|
|
964
|
+
[pricing page](https://docs.byteplus.com/en/docs/ModelArk/model-pricing)
|
|
965
|
+
publishes Seedream Pro's 2.61-million-pixel threshold; older indexed material
|
|
966
|
+
differs. The dated native model ID is retained.
|
|
967
|
+
- Google's [regular embedding prices](https://ai.google.dev/gemini-api/docs/pricing)
|
|
968
|
+
distinguish text, image, audio and video inputs. A provider that only publishes
|
|
969
|
+
some meters does not inherit the others by assumption.
|
|
970
|
+
- [BFL pricing](https://bfl.ai/pricing) distinguishes first versus additional/input
|
|
971
|
+
megapixels, ceiling rounding, pixel units and capped input sizes. Its flat
|
|
972
|
+
FLUX 1.1 image price is not fal's megapixel price. Native xAI input-image charges
|
|
973
|
+
and Alibaba's documented International-region regular rates remain qualified.
|
|
974
|
+
|
|
975
|
+
fal Tripo P2, native Tencent Hunyuan and native Bria retain verified endpoints
|
|
976
|
+
without an invented regular USD price. Bria's hosted checkpoint is not pinned by
|
|
977
|
+
its API reference. OpenRouter Muse's image-specific endpoint metadata is empty;
|
|
978
|
+
its general endpoint token meters do not establish image parameter parity.
|
|
979
|
+
These gaps are explicit qualifications, not zero-price defaults or passed paid
|
|
980
|
+
behavior tests. Real-ESRGAN's official code link is distribution metadata, not
|
|
981
|
+
an invented native hosted provider.
|
|
982
|
+
|
|
983
|
+
## Grida public service bindings (2026-10-13)
|
|
984
|
+
|
|
985
|
+
Grida bindings cover its existing 40 models and the rig-check operation; the
|
|
986
|
+
other 21 retained model cards do not claim Grida availability. Public
|
|
987
|
+
[model discovery](https://gateway.grida.ai/v1/models),
|
|
988
|
+
[endpoint discovery](https://gateway.grida.ai/v1/endpoints), and
|
|
989
|
+
[operation schemas](https://gateway.grida.ai/schemas/operations.json) establish
|
|
990
|
+
service identities and contracts. Discovery requires authentication. Synchronous
|
|
991
|
+
image, asynchronous video and specialized operation interfaces stay explicit;
|
|
992
|
+
these bindings do not imply another provider's parameter compatibility.
|
|
993
|
+
|
|
994
|
+
`published_pricing` records advertised USD per named quantity, including bands,
|
|
995
|
+
conditions and rounding. Provider-specific normalized quantities are not raw
|
|
996
|
+
image counts or guaranteed per-request totals. Source unit identifiers are
|
|
997
|
+
preserved. Grida's Gemini 3.7/3.8 Flash bindings omit prices: the inspected
|
|
998
|
+
discovery does not independently establish a regular, non-promotional schedule.
|
|
999
|
+
Another provider's regular rates are not evidence of Grida's advertised rates.
|
|
1000
|
+
Other recorded service rates match the inspected discovery.
|
|
1001
|
+
This catalogue describes public service facts without supplier selection or
|
|
1002
|
+
private financial policy.
|
package/README.md
CHANGED
|
@@ -5,7 +5,7 @@ provider bindings, and published prices, with the labs that make them and their
|
|
|
5
5
|
logos. Every fact cites a first-party source.
|
|
6
6
|
|
|
7
7
|
This package is the data, published unchanged from release
|
|
8
|
-
[`v1.
|
|
8
|
+
[`v1.20261013.1`](https://github.com/gridaco/models/releases/tag/v1.20261013.1) of [gridaco/models](https://github.com/gridaco/models). It contains no code.
|
|
9
9
|
|
|
10
10
|
| file | contents |
|
|
11
11
|
| --- | --- |
|
|
@@ -25,7 +25,7 @@ From a CDN, without installing:
|
|
|
25
25
|
|
|
26
26
|
```
|
|
27
27
|
https://cdn.jsdelivr.net/npm/@grida/ai-models@1/models.json
|
|
28
|
-
https://cdn.jsdelivr.net/npm/@grida/ai-models@1.
|
|
28
|
+
https://cdn.jsdelivr.net/npm/@grida/ai-models@1.20261013.1/models.json
|
|
29
29
|
```
|
|
30
30
|
|
|
31
31
|
## Versions
|
package/manifest.json
CHANGED
|
@@ -1,23 +1,23 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "grida-models",
|
|
3
|
-
"version": "v1.
|
|
3
|
+
"version": "v1.20261013.1",
|
|
4
4
|
"schema_version": 1,
|
|
5
|
-
"commit": "
|
|
6
|
-
"created": "2026-10-
|
|
7
|
-
"source": "https://github.com/gridaco/models/tree/
|
|
8
|
-
"digest": "sha256:
|
|
5
|
+
"commit": "e10c3c04c763dae10ab796df317b3105dcd611c4",
|
|
6
|
+
"created": "2026-10-13T12:42:45Z",
|
|
7
|
+
"source": "https://github.com/gridaco/models/tree/e10c3c04c763dae10ab796df317b3105dcd611c4/catalogue",
|
|
8
|
+
"digest": "sha256:ba54f9f0b6ce1c632064796cc443072ac7e7ec89f46f1d282bd734e9add18c50",
|
|
9
9
|
"counts": {
|
|
10
10
|
"cards": 61,
|
|
11
11
|
"labs": 14
|
|
12
12
|
},
|
|
13
13
|
"files": {
|
|
14
14
|
"PROVENANCE.md": {
|
|
15
|
-
"sha256": "
|
|
16
|
-
"bytes":
|
|
15
|
+
"sha256": "4c0b2ee977523b56abf8f706fafe167d643f9eaeed2021509daf24b789ad77e6",
|
|
16
|
+
"bytes": 44757
|
|
17
17
|
},
|
|
18
18
|
"README.md": {
|
|
19
|
-
"sha256": "
|
|
20
|
-
"bytes":
|
|
19
|
+
"sha256": "24dc2be40afb3aa64da8b04e2e664865ac6d3a7b7d5b258adb3c1241cc00d004",
|
|
20
|
+
"bytes": 16528
|
|
21
21
|
},
|
|
22
22
|
"labs.json": {
|
|
23
23
|
"sha256": "b65f0817cce0dcfec9d5393d7d8f2c5e7d13d957d71ddaf32b481f83d2f6bad4",
|
|
@@ -80,16 +80,16 @@
|
|
|
80
80
|
"bytes": 330
|
|
81
81
|
},
|
|
82
82
|
"models.json": {
|
|
83
|
-
"sha256": "
|
|
84
|
-
"bytes":
|
|
83
|
+
"sha256": "4703e28fd32e45efd792705f204c2f9de12786b5ee9eadf0c6e4d2062b2bec26",
|
|
84
|
+
"bytes": 509427
|
|
85
85
|
},
|
|
86
86
|
"schemas/labs.schema.json": {
|
|
87
87
|
"sha256": "9ddaedd56826bebdbe72b84f76cfc86df0b5c5902ba9f0259c2ea2cbb9327f99",
|
|
88
88
|
"bytes": 2469
|
|
89
89
|
},
|
|
90
90
|
"schemas/models.schema.json": {
|
|
91
|
-
"sha256": "
|
|
92
|
-
"bytes":
|
|
91
|
+
"sha256": "e7c3171e164cefb53449c50b85dcb640b57884a74e9a814980a5e3785308783c",
|
|
92
|
+
"bytes": 28929
|
|
93
93
|
}
|
|
94
94
|
}
|
|
95
95
|
}
|