ChessBench data and sources
Ranks LLMs by their ability to plan, predict and adapt.
Read the benchmark
ChessBench compares existing LLMs through chess games. Each thinking setup has its own Glicko-2 rating. Models receive the board position, piece list, accepted move history, and legal moves. Stockfish analysis is for viewers only.
ChessBench Elo ratings compare LLMs within this benchmark only. They are not equivalent to Chess.com, FIDE, or other chess ratings. Beating a 2000-rated LLM here does not establish your rating or predict the same result against a 2000-rated human player.
Live leaderboardMethodology and exact promptsPublic match records
Download current results
These files use the same saved public standings as the website. They include the result update time, retrieval time, source URLs, units, and setup sample counts. Refresh at most once per minute and use the response ETag to avoid repeated downloads.
Leaderboard JSONLeaderboard CSVLeaderboard MarkdownMethodology MarkdownAgent guide
- Rating
- Glicko-2; LLM pool only. Unranked setups have no official position.
- Cost
- USD per completed fully priced benchmark game. N/A or null means unavailable.
- Time
- milliseconds of model request time per accepted move.
- Output
- output tokens per completed rated game, including reported reasoning.
Model result pages
- Gemini 4 Argon chess results2 thinking setups
- Claude Opus 5.5 chess results2 thinking setups
- Claude Sonnet 5.5 chess results2 thinking setups
- GPT-6 Astra chess results2 thinking setups
- Claude Fable 5.1 chess results2 thinking setups
- GPT-6.1 SOL chess results2 thinking setups
- Claude Haiku 5.5 chess results2 thinking setups
- DeepSeek V4.1 Flash (1017) chess results2 thinking setups
- GPT-6.1 Terra chess results2 thinking setups
- Grok 4.7 chess results1 thinking setup
- Muse Spark 1.3 chess results2 thinking setups
- GLM 5.3 chess results2 thinking setups
- GPT-6 Luna chess results2 thinking setups
- MiniMax M3.1 Flash chess results2 thinking setups
- MiMo V2.6 Flash chess results1 thinking setup
- GLM 5.3 Flash chess results2 thinking setups
- MiniMax M3 chess results1 thinking setup
- Qwen3.7 Plus chess results1 thinking setup
Read and cite a result
ChessBench. LLM chess benchmark results. https://chessbench.org/data. Retrieved 2026-10-28. Include the model, thinking level, protocol, and sample size when citing results.
Use a model page for a setup comparison and a match page for a single result. Report the thinking level, sample size, date, and qualification state. A single game is not a general claim about a model.
Each public LLM match page links to PGN. The read-only JSON endpoint is /api/matches/{id}/export. Use actual IDs from public match pages. Administrator records and credentials are not public data.
Data use follows the current terms. No new open-data licence is granted here. Model logos and names remain their owners’ property.
Benchmark guides
How LLM chess ratings work — Read ChessBench Elo, qualification, and game results without confusing them with human chess ratings.
Thinking level, chess results, and cost — Compare reasoning settings with measured chess results, response time, output tokens, and cost per game.