API Documentation

• Read-only: GET, HEAD, and OPTIONS only. No authentication or API keys required.
• CORS is open: responses are sent with Access-Control-Allow-Origin: * so you can call the API from any client.
• Responses are cacheable for 60 seconds in browsers and one hour in shared caches, with one day of stale-while-revalidate.
• The catalog is a static, hand-curated snapshot. Every response carries meta.dataFreshness, the date the data was last refreshed.

All versioned endpoints live under /api/v1:

/api/v1

The machine-readable index at /api/v1 lists every endpoint with catalog statistics, and /api/openapi.json serves an OpenAPI 3.1 description of the API.

Every successful response is wrapped in a common envelope: data holds the payload and meta holds API version, data freshness, and endpoint-specific fields.

{
  "data": [ … ],
  "meta": {
    "apiVersion": "v1",
    "dataFreshness": "2026-09-29",
    "page": 1,
    "limit": 50,
    "total": …,
    "pageCount": …,
    "hasNextPage": true,
    "hasPreviousPage": false,
    "query": { … },
    "facets": { … }
  }
}

For list endpoints, meta also includes pagination, an echo of the applied query, and filter facets:

Errors use a parallel envelope with an error object instead of data. Validation failures return 400; unknown slugs or IDs return 404.

{
  "error": {
    "code": "NOT_FOUND",
    "message": "Model 'gpt-99' was not found.",
    "details": { "resource": "model", "value": "gpt-99" }
  },
  "meta": {
    "apiVersion": "v1",
    "dataFreshness": "2026-09-29"
  }
}

Paths below are relative to /api/v1. Braces mark path parameters.

/models and /releases accept:

/image-models and /video-models accept a smaller set:

# Five highest-scoring models on SWE-bench Verified
GET /api/v1/models?benchmark=swe-bench-verified&limit=5

# Open-weight Google models, largest context window first
GET /api/v1/models?organization=Google&openSource=true&sort=contextWindow

# One model, with same-family related models in meta
GET /api/v1/models/claude-opus-5

# Compare two models, or address the comparison by canonical slug
GET /api/v1/compare?a=claude-opus-5&b=gpt-5-6-sol
GET /api/v1/compare/claude-opus-5-vs-gpt-5-6-sol

The same calls work from a local dev server with curl, for example:

curl -s "http://localhost:3000/api/v1/models?benchmark=swe-bench-verified&limit=5"

• Values that cannot be verified from a trustworthy source are omitted rather than estimated. Treat absent fields as unknown, never as zero.
• Benchmark scores keep each benchmark's native metric and version, so scores with the same name are not always comparable across benchmark variants. Check the benchmark metadata returned by /benchmarks before comparing.
• Some coding benchmarks evaluate an agent system (scaffold, tools, harness) rather than a raw model; the benchmark metadata notes this where relevant.
• Prices and specifications are point-in-time observations. Verify critical numbers with the provider before relying on them.