Can Claude play chess?

Yes, Claude can play chess. It knows the rules, the openings and the main ideas, and the newest Claude models play a legal move almost every time. But Claude is a language model, not a chess engine. It doesn't calculate long lines of moves the way Stockfish does, so it still blunders, and older or faster Claude models can lose track of the board and try moves that aren't allowed.

Here is what public tests show, as of October 22, 2026. The numbers come from the AI Chess Leaderboard run by the developer Dubesor (data updated October 21, 2026), which records every move its AI players make.

On that leaderboard a blunder is a move that cuts the player's chances of winning by 15% or more.

Can Claude play chess well?

Better than most people expect, and far below a real engine. On Dubesor's leaderboard, Claude Opus 5.5 has an Elo of 1571 ±97 after 35 games. In the same pool, maia5-1500, a neural network trained on more than 10 million games by 1500-rated Lichess players, sits at 1554 ±90. A bot that plays random legal moves sits at 419. On Maxim Saplin's LLM Chess Leaderboard, which ties its scale to the Komodo Dragon engine's levels on Chess.com, Claude Opus 5.5 with high reasoning scores 1575 ±113 (data of October 20, 2026).

Both authors warn that these numbers only mean something inside their own leaderboard. They are not Chess.com or FIDE ratings. A fair reading: in these tests, the best Claude plays roughly as well as a human-like 1500 bot, with wide error bars. We compare every leaderboard, and explain why they disagree, in What is Claude's chess Elo?

Earlier Claude models were much weaker. On the same board, Claude 3.5 Sonnet is at 600 ±93 and Claude Sonnet 4.5 at 659 ±72. So "Claude is bad at chess" was a fair complaint a year or two ago.

Why Claude is sometimes bad at chess

A real game shows it. In August 2025, Kaggle and Google DeepMind ran an AI chess exhibition between eight language models. Claude Opus 4 lost its quarterfinal 4–0 to Gemini 2.5 Pro, according to Chess.com's report of August 5, 2025. In the first game Claude, playing Black, had a normal position after nine moves. Then it pushed a pawn in front of its own castled king (g5).

r . b q . r k .
p p . . b p p .
. . n p p n . p
. . . . . . . .
. . . N P P . B
. . N . . . . .
P P P Q . . P P
. . K R . B . R

Castling tucks the king in a corner behind a wall of pawns, and those pawns are its shelter. In our check with Stockfish, the position before the pawn push is about equal (around +0.4 for White). After it, White takes the pawn with a double attack, one move that threatens two things at once, and the evaluation jumps to close to +3 for White. Chess.com's annotator notes that Claude's own reasoning saw the weakness and judged it acceptable to get counterplay. It wasn't. Chess.com also noted that this was the only quarterfinal with more games ending in checkmate than in forfeits for illegal moves.

Three things cause mistakes like this:

• Claude predicts moves; it doesn't search them. An engine checks the replies to a move, then the replies to those, many moves deep. A language model writes the move that seems most likely and gives reasons for it.
• It has to keep the board in its head. In a chat, the board is only text. That's why Claude Haiku 4.5's legal moves drop from 98.6% to 51.6% when it sees only the move list. In the Game Arena paper by Google DeepMind and Kaggle (September 2026), models got four tries to play a legal move before losing the game, and Claude Haiku 4.5 often needed more than ten retries to finish one.
• Confidence isn't accuracy. This is true of every chatbot. In June 2025, ChatGPT lost to the 1979 Atari 2600 Video Chess cartridge after mixing up its pieces, as IBM Think reported. Claude wasn't in that test; see Can ChatGPT play chess? and why chatbots play illegal moves.

Can Claude Opus play chess better than Sonnet or Haiku?

Yes, on today's leaderboards. Dubesor's board, as of October 21, 2026:

• Claude Opus 5.5: 1571 ±97
• Claude Fable 5: 1465 ±88
• Claude Sonnet 5.5: 1321 ±108
• Claude Haiku 5.5: 1068 ±102

Letting the model think before it moves also matters a lot. On the open LLM Chess Benchmark, Claude Opus 4.5 is rated 1012 with high thinking and 175 with thinking off. If you want to try a game yourself, pick the strongest model you have and turn on extended thinking.

How to play a better game against Claude

• Share the position every turn. Paste the full move list or the position code (FEN) with each move. On Dubesor's leaderboard, Claude models given the position and its legal moves make far fewer illegal moves than when they only see the moves played.
• Keep a real board open. Play the moves on Lichess or Chess.com's analysis board so you can catch an illegal move right away.
• Don't trust its evaluations. When Claude says a position is winning, that's its opinion, not a calculation.

Our guide to playing chess with Claude walks through the full setup.

How Chess Coach helps

Chess Coach doesn't make Claude a stronger player, and it won't play against you. It does the job a language model can't: checking the facts. It's a connector for Claude that reviews your finished Chess.com and Lichess games with Stockfish on our own server. Claude then explains the key moments in plain words on an interactive board, with a red arrow for the move you played and a green arrow for a better one. It never helps with games in progress, and the free plan includes one in-depth review a day. Connect Chess Coach to Claude, see how the AI chess coach works or read the docs.