Public API

Free, read-only JSON over the official rating snapshots. No authentication, no keys. Version 1 lives under /api/v1/; additive fields won't break you, breaking changes would ship as /api/v2 alongside.

Ground rules

Rate limit: per-IP, fixed windows. Exceeding it returns 429 with a Retry-After header and body {"code":"RATE_LIMITED"} — back off and retry.
Caching: responses carry Cache-Control: public, max-age=300. Data derives from snapshots computed every 6 hours — polling more often than a few minutes gains nothing.
Bulk analysis? Don't crawl the API — the nightly dumps contain every eligible battle and vote.
Unknown categories/models return 404. All timestamps are ISO 8601 UTC.

GET/api/v1/categories

All active battle categories with the timestamp of their latest official rating snapshot.

Live example: /api/v1/categories

GET/api/v1/leaderboards/{category}

The latest official leaderboard of a category: every rated model with raw AND style-controlled ratings, 95% bootstrap confidence intervals, vote counts, provisional flags, and the published rank bands (overlapping CIs share a rank; provisional models are rated but unranked).

category (path) — category slug from /api/v1/categories (e.g. human-like-chat)

computed_at is null and entries is empty while a category has no snapshot yet.

Field names mirror the published leaderboard.json (DUMP_FORMAT.md §6).

Live example: /api/v1/leaderboards/human-like-chat

GET/api/v1/models

Every listed model with public metadata: provider, license, open-weights flag, context window, list pricing, release date, tier, mock flag, plus model_type, input_modalities, output_modalities and capabilities.

Listed = in the arena roster, or carrying official ratings or curated benchmark scores. Models we have not evaluated ourselves yet appear too — their arena ratings are simply absent.

Live example: /api/v1/models

GET/api/v1/models/{slug}

One model's profile: metadata plus its latest official rating (both variants) in every category it is rated in, and its latest curated third-party benchmark score per benchmark (benchmarks[]) with full provenance.

slug (path) — model slug from /api/v1/models (e.g. claude-fable-5-1)

Live example: /api/v1/models/claude-fable-5-1

GET/api/v1/models/{slug}/history

The full rating history of a model: every official snapshot row (raw and style-controlled variants), oldest first per category.

slug (path) — model slug
category (query (optional)) — restrict to one category slug

Live example: /api/v1/models/claude-fable-5-1/history?category=human-like-chat

GET/api/v1/benchmarks

The curated third-party benchmark catalogue: active sources (with their terms notes), the metric-group taxonomy, and every active benchmark with unit, direction, scale, description, methodology link, listed-model coverage and latest observation date.

External scores are curated imports (content files + admin entry — no scrapers), shown as context with provenance; our blind-vote arena ratings remain the primary signal.

Live example: /api/v1/benchmarks

GET/api/v1/benchmarks/{slug}

One benchmark with the latest score per listed model, best value first. Every score row carries value, ci_low/ci_high, and its provenance: observed_at, retrieved_at, source_url and source_version.

slug (path) — benchmark slug from /api/v1/benchmarks (e.g. lmarena-text)

Unknown or inactive benchmarks return 404.

Live example: /api/v1/benchmarks/lmarena-text

GET/api/v1/boards/{modelType}

The unified board of one model type: every listed model with its metadata, capabilities and humaneval_evaluated flag, our arena ratings per category (rank bands match /api/v1/leaderboards), telemetry medians from our own battles, and the latest external score per benchmark of that type.

modelType (path) — one of llm, image-generation, video-generation, text-to-speech, speech-to-text, other

Live example: /api/v1/boards/llm

Example

curl -s https://humaneval.org/api/v1/leaderboards/human-like-chat | jq '.entries[] | {model, rating, rank}'