Model Capability Registry — References

Sources for the facts an entry records, and for the schema conventions used to record them.

Provider-published capability and pricing surfaces

• OpenAI models documentation — context windows, maximum output, modalities, and feature support per model identifier.
• Anthropic models overview — context windows, output limits, and per-model feature availability.
• Google Gemini API documentation — model variants, input modalities, and token limits.

Provider list-models API endpoints return the identifiers an account can actually call, which is the only authoritative answer to availability under a specific tier and region; they do not return pricing.

Open-weight model metadata

• Hugging Face Hub model cards documentation — the card format, including the structured metadata block carrying licence and task tags.
• Hugging Face Transformers configuration documentation — config.json fields such as maximum position embeddings, the ground truth for context length on self-hosted weights.

Schema and record keeping

• JSON Schema specification — for validating registry entries in CI so an unqueryable free-text field cannot be committed.
• OpenTelemetry semantic conventions — defines the attribute names for model identity and token counts, worth aligning registry field names with so telemetry and registry join cleanly.

Pricing pages are deliberately not linked: they change without notice and a stale link here would be worse than no link. Treat the provider's current pricing page as the only source.