Can ChatGPT play chess?

ChatGPT can play chess in the sense that it will name a move every turn, and it knows the rules, the openings and the vocabulary. It can't play it well. In June 2025 ChatGPT lost to a 1979 Atari 2600 cartridge on its easiest level (The Register, 9 June 2025), and in published tests chatbots still try moves the rules don't allow, especially deep into a game.

What the tests show

There's no official "ChatGPT chess rating". What exists is a handful of public tests, each with a different model and a different method. Here they are in date order, as of October 2026:

Date	Test	Model	Result
Sep 2023	Mathieu Acher's games against Stockfish 16	gpt-3.5-turbo-instruct (a text-completion model, not the chat one)	About 1750 Elo; an illegal move in 16% of games
Sep 2023	Same study	GPT-4 (chat)	About 1371 Elo counting only legal games; an illegal move in 32% of games
Nov 2024	Dynomight vs Stockfish on its lowest level	GPT-4o, GPT-4o mini, o1-mini	All rated "terrible"; only gpt-3.5-turbo-instruct played well
Jun 2025	Robert Caruso vs Atari 2600 Video Chess	ChatGPT (GPT-4o)	Lost on the beginner setting after about 90 minutes
Jul 2025	Magnus Carlsen, a casual online game	ChatGPT	Carlsen won in 53 moves without losing a piece
Aug 2025	Kaggle Game Arena exhibition	o3	Won the bracket, beating Grok 4 4–0 in the final
Sep 2026	Game Arena paper, chess leaderboard	o3, GPT-5.2	1009 and 933 on the arena's internal scale (top model: Gemini 3 Pro Preview, 1325)

Elo is the rating system chess uses: the higher the number, the stronger the player. Keep in mind that the numbers above come from different pools and can't be lined up against each other.

Can ChatGPT play chess well now?

Better than in 2023, still not well. Three things stand out from the table.

The chat version was the weak one. The 1750 estimate that gets quoted online belongs to gpt-3.5-turbo-instruct, an older model that completes text rather than chats. In the same study the chat-tuned GPT-4 scored far lower and made illegal moves in almost a third of its games, and the chat-tuned gpt-3.5-turbo made them in 93%. Acher called this the first solid evidence that training for chat made the models worse at chess. A year later Dynomight found the same pattern with GPT-4o.

Reasoning models closed some of the gap. OpenAI's o3, a model that "thinks" before answering, won all 12 of its games at Kaggle's August 2025 exhibition, although the first four came by forfeit: its opponent, Kimi k2, couldn't produce a legal move in four tries (Chess.com, 5 August 2025). It then won the semifinal and the final 4–0 each. On the arena's full leaderboard, published in September 2026, o3 sits at 1009 and GPT-5.2 at 933. Those numbers rank AI models against each other, with DeepSeek V3.2 fixed at zero; they are not a human rating like your Chess.com number.

The gap to an engine is still huge. Stockfish 19 is rated 3650 on the CCRL 40/15 engine list computed on 16 October 2026. Even after a big jump in quality, no chatbot is in that conversation. We cover the head-to-head in ChatGPT vs Stockfish.

Does ChatGPT understand chess?

It understands chess the way it understands everything: as text. It learned from a huge amount of chess writing, from the rules and game records to opening names and commentary. Ask it what a pin is, or why you castle early, and you'll get a good answer.

What it doesn't have is a board. Each turn it has to rebuild the position from the moves typed so far, and nothing inside it checks that its move is allowed. Caruso watched ChatGPT confuse rooks with bishops, miss simple pawn forks and repeatedly lose track of where the pieces were. Even when Kaggle handed the models the exact position every turn, illegal attempts clustered in the later part of games, once the memorized openings ran out.

There's a funny footnote. After Carlsen beat it, ChatGPT estimated his strength at around 1800 to 2000, according to the screenshots he shared and TIME's report (17 July 2025). He is the world's top-rated player. We explain the mechanics in why ChatGPT plays illegal moves.

Can ChatGPT play chess for you, or with you?

With you, yes, as a game for fun:

• Keep the real game on a board of your own, such as a free analysis board on Lichess.
• Paste the full move list or the position each turn, not just your last move.
• When it suggests something impossible, tell it the move is illegal and give the position again.

For you, no. If you mean "suggest moves in my online games", that's cheating. Lichess bans any external help during a game, and Chess.com bans software of any kind. It would also play badly.

Where ChatGPT is genuinely useful is after the game: explaining an idea, quizzing you on a plan, or turning advice into a study routine. Our ChatGPT chess coach prompt guide shows how. Just don't trust it to evaluate a position on its own; see can ChatGPT analyze chess games.

The better split: an engine plays, an assistant explains

The honest fix isn't a smarter prompt. It's giving each job to the right tool. A chess engine like Stockfish finds the moves and the evaluations; a language model turns them into words a beginner can follow.

That's how Chess Coach works. It's a connector for Claude and ChatGPT: it reads your finished Chess.com or Lichess games, runs Stockfish on our own server, and your assistant explains the key moments on an interactive board, with arrows for the move you played and the better one. Every evaluation comes from the engine, not from the model's guess. The free plan includes one in-depth review a day. In ChatGPT, install Chess Coach from OpenAI’s plugin directory; see can ChatGPT analyze chess games?. To try it, see how to connect Chess Coach or the docs.