model-metrics-api

Scrapes model metrics from artificialanalysis.ai and serves them via a REST API and MCP server.

task build
bin/scrape-aa serve

# All models, sorted by Intelligence Index (default)
curl http://localhost:8080/api/v1/models | jq '.models[0:5] | .[].name'

# Anthropic models sorted by coding score
curl "http://localhost:8080/api/v1/models?creator=anthropic&sort_by=coding_index" | jq '.models[].name'

# Models scoring above 0.5 on GPQA, cheapest first
curl "http://localhost:8080/api/v1/models?bench=gpqa&min=0.5&sort_by=price_input&order=asc" \
  | jq '.models[] | {name, gpqa, price1mInputTokens}'

# Single model
curl http://localhost:8080/api/v1/models/claude-sonnet-4-5 \
  | jq '{name, intelligenceIndex, price1mInputTokens}'

Scrape — one fetch from artificialanalysis.ai, timestamped and stored in SQLite. The server scrapes on startup and then on each --interval tick. Every scrape is kept, so you can query historical data.

Slug — a stable URL-safe identifier for each model (e.g. claude-sonnet-4-5). Slugs are consistent across scrapes, so you can track a model over time.

bin/scrape-aa scrape [--out models.json]

Fetches all models, writes a JSON file, prints the top 5 by Intelligence Index.

bin/scrape-aa serve [--addr :8080] [--db ./data.db] [--interval 1h]

Scrapes immediately on start, then on each interval. Stores history in SQLite.

Base: http://localhost:8080/api/v1

All parameters are optional and compose freely.

Examples:

# Models from a specific historical scrape
curl "http://localhost:8080/api/v1/models?scrape_id=3"

# Models as of a specific point in time
curl "http://localhost:8080/api/v1/models?as_of=2025-01-01T00:00:00Z"

# Combine filters and sort
curl "http://localhost:8080/api/v1/models?creator=openai&bench=hle&min=0.2&sort_by=price_input&order=asc"

# Historical model snapshot
curl "http://localhost:8080/api/v1/models/gpt-4o?scrape_id=1"

Pass one of these field names as the bench query parameter.

Scrape object (returned by /scrapes and POST /scrapes):

{
  "id": 4,
  "scrapedAt": "2025-05-31T12:00:00Z",
  "modelCount": 183
}

Models response (returned by GET /models):

{
  "scrapeId": 4,
  "scrapedAt": "2025-05-31T12:00:00Z",
  "models": [ ...model objects... ]
}

Model object — key fields:

{
  "id": "claude-sonnet-4-5",
  "slug": "claude-sonnet-4-5",
  "name": "Claude Sonnet 4.5",
  "shortName": "Sonnet 4.5",
  "modelCreatorName": "Anthropic",
  "modelCreatorSlug": "anthropic",

  "intelligenceIndex": 76.4,
  "codingIndex": 71.2,
  "agenticIndex": 68.0,
  "gpqa": 0.718,
  "hle": 0.241,

  "price1mInputTokens": 3.0,
  "price1mOutputTokens": 15.0,
  "cacheHitPrice": 0.3,

  "contextWindowTokens": 200000,
  "totalParameters": null,
  "reasoningModel": false,
  "isOpenWeights": false,
  "deprecated": false,

  "medianOutputTokensPerSecond": 98.4,
  "medianTimeToFirstTokenSeconds": 0.61,

  "inputModalityText": true,
  "inputModalityImage": true,
  "inputModalitySpeech": false,
  "outputModalityText": true
}

Nullable fields are null when the data isn't available for that model. The full model object contains ~60 additional fields for pricing blends, token-count breakdowns, openness metadata, and per-domain omniscience scores.

SSE endpoint: http://localhost:8080/mcp/sse

Connect with any MCP-compatible client (Claude Desktop, Continue, etc.) using the SSE transport.

Returns models from the latest scrape as a JSON array.

Returns a single model by slug.

Returns "model not found" if the slug doesn't exist.

Case-insensitive substring search across model name and creator name.

task build     # compile to bin/scrape-aa
task test      # go test -race ./...
task lint      # golangci-lint + govulncheck + nilaway + go-arch-lint
task generate  # sqlc generate (after schema/query changes)

Requires

• Task
• golangci-lint
• govulncheck
• nilaway
• sqlc
• go-arch-lint