Modelsheet is an open-source capability registry for generative media models. It answers two practical questions: what can this model actually produce, and what does that output cost from the lab that made it?
Live registry: aj-base44.github.io/modelsheet
The registry is media-shaped: image sizes, video durations, reference inputs, generated audio, delivery formats, and the pricing units that apply to those capabilities. It is not an LLM registry, does not compare token context windows or reasoning features, and does not include reseller or aggregator pricing.
The differentiator is a drift feed derived automatically from the Git history of the model records. It shows the exact structured field that changed, its before and after values, the affected model, commit, and date. Contributors never maintain a second changelog.
Coverage: 53 total models: 52 documented, 1 indexed (18 image, 18 video, 17 audio).
An indexed record establishes the official model identity and first-party source links but does not claim comparison-ready capability data. A documented record contains at least one real, sourced capability constraint and non-unknown lab pricing, so it can appear in capability filters and comparisons.
Each model has one TOML file at models/<lab>/<model>.toml. The model is the primary entity; API identifiers are access metadata for that same model, not separate provider endpoints.
A record contains:
- official model identity and any sourced colloquial aliases;
- task-specific capability profiles, so combinations such as duration, size, aspect ratio, and reference inputs are not falsely cross-multiplied;
- a flat list of lab-published pricing charges;
- explicit conflicts when official sources disagree; and
- a source ledger with the exact official URL and retrieval date behind each claim.
unknown, unsupported, and not_applicable are intentionally different. Missing documentation is never converted to false or an empty list. See SCHEMA.md for the complete contract.
verified: a human compared the record's stated values with the cited live first-party pages. Agents and automated jobs must never assign this state.unverified: claims have official source references, but a human has not completed that comparison. This is the default for contributions and for every new record created by an automated research run.
Older records without a [verification] table have no recorded verification state. Consumers must display that as not recorded, never infer verified, and never treat a citation as a substitute for human verification.
Contributors edit TOML only. Tooling validates the source records and builds a deterministic artifacts/api.json for consumers. The static site reads that artifact; there is no database or backend.
Public machine-readable outputs:
- latest registry JSON and version-pinned v1 JSON;
- latest semantic drift JSON and version-pinned v1 drift JSON;
- RSS drift feed; and
- source v1 JSON Schema.
The unversioned URLs follow the latest compatible contract. Consumers that need a fixed shape should use the version-pinned URLs; the policy is recorded in decision 0002.
The repository also includes a typed, dependency-free package at packages/registry with ESM, CommonJS, raw JSON, and TypeScript exports. @modelsheet/registry is publish-ready but intentionally not published until a maintainer chooses a release cadence. The offline modelsheet agent skill can query documented models, constraints, prices, and provenance directly from the local artifact.
After installing Node.js 22 or newer:
npm.cmd install
npm.cmd ci --prefix packages/registry
npm.cmd ci --prefix site
npm.cmd run check
npm.cmd run package:check
npm.cmd run check --prefix sitenpm.cmd run build creates deterministic registry, drift JSON, and RSS artifacts. It requires full Git history and fails loudly in a shallow clone rather than publishing a partial feed.
Read CONTRIBUTING.md and SCHEMA.md before opening a pull request. Use only official lab documentation, schemas, pricing pages, changelogs, or release notes. Every factual claim needs a source URL and the date it was read. When in doubt, record less and use unknown.
Modelsheet is released under the MIT License.
The static site is deployed through GitHub Pages at the repository URL above. No custom domain, CNAME, DNS configuration, backend, or runtime data lookup is required.