What is Claude's chess Elo?
Claude has no official chess Elo. Claude doesn't hold an official Chess.com, Lichess or FIDE rating, and we couldn't find a chess rating published by Anthropic. What exists are independent LLM chess leaderboards that make AI models play each other or a chess engine. As of October 22, 2026, the strongest Claude on them, Claude Opus 5.5, is rated about 1570 on two of them. Those numbers only make sense inside each leaderboard.
Claude chess Elo on four leaderboards
| Leaderboard | How it rates | Best Claude result | For comparison |
|---|---|---|---|
| AI Chess Leaderboard by Dubesor, updated Oct 21, 2026 | Elo from games between AI models, relative to its own pool | Claude Opus 5.5: 1571 ±97 (35 games) | Leader: Gemini 3.1 Pro Preview, 1860 ±85. Human-like maia5-1500 bot: 1554 ±90. Random legal moves: 419 |
| LLM Chess Leaderboard by Maxim Saplin, data of Oct 20, 2026 | Games against a random player and the Komodo Dragon engine, whose levels are rated on Chess.com | Claude Opus 5.5 (high reasoning): 1575 ±113 (64 games) | Leader: GPT-6 Astra (high), 1614 ±110 |
| LLM Chess Benchmark, checked Oct 22, 2026 | Glicko-2 rating against fixed engine and bot anchors, as an estimated Lichess Classical rating | Claude Opus 5.5: 1197 from only 2 games (95% range 517–1876). Claude Opus 4.5 (high): 1012 from 17 games | Top AI model: Gemini 3.1 Pro Preview (high), 2144 |
| Game Arena paper by Google DeepMind and Kaggle, Sept 2026 | Ratings between models only; the weakest model is set to 0 | Claude Opus 4.5: 236 (95% range 163–323) | Leader: Gemini 3 Pro Preview, 1325. DeepSeek V3.2: 0 |
Same company, four very different numbers. None of them is wrong. They measure different things, with different opponents and different rules. Kaggle also keeps a live chess leaderboard that changes as new matches are played; we didn't copy its numbers here.
Opus, Sonnet, Haiku: how the Claude models compare
Saplin's leaderboard has tested most Claude versions with the same method, so it shows the trend well. Ratings as of October 20, 2026:
| Claude model (setting) | Elo | Games |
|---|---|---|
| Claude Opus 5.5 (high reasoning) | 1575 ±113 | 64 |
| Claude Opus 5 (adaptive thinking, high) | 1285 ±127 | 53 |
| Claude Sonnet 5.5 (high reasoning) | 1195 ±120 | 74 |
| Claude Haiku 5.5 (high reasoning) | 1143 ±108 | 60 |
| Claude Sonnet 5 (high reasoning) | 806 ±79 | 134 |
| Claude Opus 4.5 (16,000-token thinking budget) | 446 ±141 | 49 |
| Claude Opus 4.5 (default settings) | 71 ±96 | 66 |
Two patterns stand out. Newer generations are much stronger: Opus 5.5 is more than 1,000 points above Opus 4.5 on this scale. And within one model, the settings change the result a lot: Opus 4.5 gains almost 400 points when it gets a thinking budget. Saplin hasn't listed Claude Fable; on Dubesor's board, Claude Fable 5 is at 1465 ±88, below Claude Opus 5.5.
Why LLM chess Elo is so hard to measure
Elo is a simple idea: beat stronger players and your number goes up. For language models, five things get in the way.
- Elo only works inside one pool. A rating says how you score against the players you actually met. Dubesor's page says it plainly: Elo is always relative to the pool it was measured in. Saplin notes his numbers can't be compared across pools either. The Game Arena paper goes further and sets its weakest model to 0, which is why Claude Opus 4.5 shows 236 there.
- Few games mean wide error bars. Claude Opus 5.5 has played only two games on the LLM Chess Benchmark, so its rating could be anywhere from 517 to 1876. Its 1197 there tells you very little.
- Settings change the player. On that same benchmark, Claude Opus 4.5 is rated 1012 with high thinking and 175 with thinking off. The prompt matters too: Dynomight showed in November 2024 that simple tricks, like asking the model to repeat the whole game before each move, made GPT-4o play much better. Claude wasn't part of that test.
- Illegal moves count as losses. In the Game Arena, a model gets four tries to play a legal move, then loses the game. Saplin's harness scores a loss if a model makes three mistakes, such as an illegal move or a badly formatted answer, while choosing a single move. So part of each rating measures following instructions, not chess.
- Models change every few months. In August 2025, Claude Opus 4 lost 4–0 to Gemini 2.5 Pro in Kaggle's AI chess exhibition, as Chess.com reported. A year later, Claude Opus 5.5 sits within the margin of error of the leader on Saplin's board. A rating describes one version, with one setting, at one point in time.
How strong is an LLM chess Elo of 1570?
Respectable for a language model, and nowhere near an engine. On Dubesor's board, Claude Opus 5.5 is level with maia5-1500, a network trained on more than 10 million games by 1500-rated Lichess players. Dubesor adds a rough, low-confidence comparison from his own tests: about 1850 on his board (the best AI models) matched Chess.com's 2000-rated bots, and the leader, Gemini 3.1 Pro Preview, lost to bots rated 2100, 2200 and 2400.
Chess engines aren't on these leaderboards as rivals at all. They're the referee: Dubesor judges the AI players' moves with Stockfish 17.1. For a head-to-head look, see ChatGPT vs Stockfish.
Why Claude's Elo matters less than you think
If you want to get better at chess, you don't need Claude to be a strong player. You need correct facts about your own moves, and someone patient to explain them. Chess Coach splits that job. Stockfish runs on our server and analyzes your finished Chess.com or Lichess games; Claude turns the engine's findings into plain words on an interactive board. Every evaluation you see comes from the engine, so Claude's own rating doesn't limit the quality of the review. See how the analysis works, connect Chess Coach or read Can Claude play chess? for what Claude does well on its own.
Frequently asked questions
What is Claude Opus's chess Elo?
As of October 2026, Claude Opus 5.5 is rated 1575 ±113 on Maxim Saplin's LLM Chess Leaderboard and 1571 ±97 on Dubesor's AI Chess Leaderboard. Both are estimates inside each leaderboard's own pool, not Chess.com or FIDE ratings.
Is there an LLM chess leaderboard?
Several. The best known are Dubesor's AI Chess Leaderboard, Maxim Saplin's LLM Chess Leaderboard, the open LLM Chess Benchmark and Kaggle's Game Arena chess leaderboard. Each uses its own rules, so their Elo numbers can't be compared directly.
Which LLM has the highest chess Elo?
It depends on the leaderboard. As of October 2026, Gemini 3.1 Pro Preview leads Dubesor's board at 1860 ±85, while GPT-6 Astra leads Saplin's at 1614 ±110, with several models, Claude Opus 5.5 among them, inside its margin of error.
What is Claude Fable's chess Elo?
On Dubesor's AI Chess Leaderboard, Claude Fable 5 is rated 1465 ±88 and Claude Fable 5.1 1305 ±100, as of October 21, 2026. That puts both below Claude Opus 5.5 on the same board.
Is an LLM chess Elo the same as a Chess.com rating?
No. An LLM's Elo is measured against other AI players under one leaderboard's rules. The leaderboards' own authors say their numbers aren't comparable with Chess.com, Lichess or FIDE ratings.