THE ARENA LEADERBOARD DATASET
Find your way through the rankings.
Explore a snapshot. Search a model. See how it changed.
Model history
Published snapshots only. Gaps mean no published score. Methodology can change over time.
Leaderboard
| Compare | ||||||
|---|---|---|---|---|---|---|
| Loading leaderboard… | ||||||
How to read this data
Arena score is the published rating for the selected arena and category. The range below each score is its published confidence interval. Missing values appear as a dash.
OpenRouter prices are current catalog rates in USD per million tokens, including on historical leaderboards. Input and output are billed separately. Cached input and long-context tiers appear when published. Image output is labeled in image tokens. Request fees are separate; image, audio, search, and other usage can add charges. Provider rates and the final cost can vary.
Base model means the Arena entry includes settings such as reasoning effort or resolution; the displayed catalog model supplies its rates and capabilities. The actual catalog name is visible on the row. A catalog alias shows its resolved release. Models without a reliable match show “No verified OpenRouter match.” Zero text rates on media models are not treated as free generation.
Context and max output are rounded token limits published by OpenRouter. Capabilities indicate supported inputs and API features. They apply to the matched catalog model. Prices are refreshed when the published site is rebuilt; the checked time is shown above the table.
Style adjusted accounts for response style; it is the default for text and vision. Raw shows the unadjusted leaderboard. Agents use IPS scores, a different scale. Scores from different arenas or scoring methods should be read separately.
Rank is the rank supplied by the dataset. Ties and gaps may occur. Votes and observations are counts per model; adding them does not give a count of unique battles or sessions.
Snapshots are published dates for the selected category. The newest date can differ by category and arena. A model’s history uses its exact published name and ends at your selected snapshot. Names are not merged across releases.
The rating methodology changed over time: Bradley–Terry in January 2024, the text and vision style adjustment and offset in May 2025, and frequency reweighting in July 2025. Large historical changes may reflect methodology changes.