Quality and attribution

Each task ranks models by a public benchmark first (for example LMArena for chat, Epoch AI for reasoning, MTEB for embeddings), then a fallback source, then a curated tier. The answer's quality.basis names the source and signal, and quality.kind says whether it was measured or estimated.

Scores run from 0 to 1 relative to the leader of that benchmark. Arena ratings within the leader's confidence interval count as 1.

Attribution

Where you show a score to people, show its quality.attribution with it: some sources are licensed CC-BY. The sources are:

• LMArena leaderboard dataset (CC-BY-4.0)
• Epoch AI Benchmarking Hub (CC-BY-4.0)
• MTEB results (CC0-1.0)
• models.dev (MIT)
• LiteLLM model prices and context windows (MIT)
• Portkey pricing configs (MIT)
• ECB euro foreign exchange reference rates (ECB reuse terms (source acknowledged))
• Nanner Tech curated tiers (Nanner Tech)

GET /v1/licenses returns the same list, and the attributions page links each source.