# Claude vs ChatGPT at chess: which AI plays better?

> Claude vs ChatGPT at chess: OpenAI's best model is slightly ahead of Claude Opus 5.5 on public LLM leaderboards, and Gemini leads both. October 2026 data.

- URL: https://thechesscoach.netlify.app/answers/claude-vs-chatgpt-chess/
- Updated: 2026-10-25
- Author: Ignacio Soler

Claude vs ChatGPT at chess is close, and neither is the best chatbot at it. As of October 25, 2026,
OpenAI's strongest tested model, GPT-6 Astra, sits slightly above Claude Opus 5.5 on two public LLM
chess leaderboards, with overlapping error bars, and Google's Gemini leads both. None of them plays
anywhere near a real chess engine like Stockfish.

One caveat before the numbers. "ChatGPT" is an app, not a single model. The leaderboards test OpenAI
models directly, with a fixed amount of "thinking" time, and the model you get in ChatGPT depends on
your plan and settings. The same is true for Claude: Opus, Sonnet and Haiku play very differently.

Here is the best model from each company on three public tests. Elo is the rating system chess uses:
higher means stronger. Each number only means something inside its own leaderboard.

| AI family | [AI Chess Leaderboard](https://dubesor.de/chess/chess-leaderboard) (Dubesor, Oct 25, 2026) | [LLM Chess Leaderboard](https://maxim-saplin.github.io/llm_chess/) (Saplin, data of Oct 20, 2026) | [Game Arena paper](https://arxiv.org/abs/2609.31473) (Kaggle, Sept 2026) |
|---|---|---|---|
| Gemini (Google) | Gemini 3.1 Pro Preview: 1860 ±85 | Gemini 4 Argon (high): 1609 ±131 | Gemini 3 Pro Preview: 1325 |
| ChatGPT models (OpenAI) | GPT-6 Astra: 1637 ±86 | GPT-6 Astra (high): 1614 ±110 | o3: 1009 |
| Claude (Anthropic) | Claude Opus 5.5: 1571 ±97 | Claude Opus 5.5 (high): 1575 ±113 | Claude Opus 4.5: 236 |
| DeepSeek | DeepSeek V4 Flash (0731): 1438 ±99 | DeepSeek R1 (0528): 300 ±134 | DeepSeek V3.2: 0 (the scale's zero point) |
| Grok (xAI) | Grok 4: 1415 ±118 (retired) | Grok 4.7 (high): 890 ±217 | Grok 4: 773 |

For scale: on Dubesor's board, maia5-1500, a human-like network trained on games by 1500-rated Lichess
players, sits at 1554 ±90, and a bot that plays random legal moves sits at 419. The Game Arena paper
tested older models and sets its weakest one to 0, so its numbers are much lower and can't be compared
with the other two columns.

## Claude vs ChatGPT chess: the head-to-head numbers

On today's models the gap is small. On Saplin's leaderboard, GPT-6 Astra is at 1614 ±110 and Claude
Opus 5.5 at 1575 ±113, well inside each other's margin of error. On Dubesor's board the gap is about
65 points, again with overlapping ranges. Claude Opus 5.5 also blunders a little less there: 2.8% of its
moves lost a big chunk of its winning chances, against 3.3% for GPT-6 Astra. Dubesor counts a move as a
blunder when it cuts the player's chances of winning by 15% or more.

A year ago the picture was different. In the Game Arena paper, which rates models tested through 2025
and early 2026, OpenAI's o3 scores 1009 and Claude Opus 4.5 only 236. So "ChatGPT is better at chess
than Claude" was a fair summary until recently. With the newest generation, it's roughly a tie.

Settings matter more than the brand. On the open [LLM Chess Benchmark](https://chessbenchllm.onrender.com/)
(checked October 25, 2026), Claude Opus 4.5 is rated 1012 with high thinking and 175 with thinking off,
and GPT-5.2 is rated 1039 with high thinking and 645 with none. A fast, non-thinking chat model of
either brand plays far worse than the same company's best model given time to think. That benchmark has
only two games for Claude Opus 5.5 so far, so we left it out of the table.

## Claude vs Gemini, and why Gemini leads

Gemini is the family to beat. It tops Dubesor's board and the Game Arena paper, and it leads the LLM
Chess Benchmark, where Gemini 3.1 Pro Preview (high) is at 2144. In Kaggle's first AI chess exhibition,
in August 2025, Gemini 2.5 Pro beat Claude Opus 4 4–0 in the quarterfinals, as
[Chess.com reported](https://www.chess.com/news/view/kaggle-game-arena-chess-2025-day-1).

Part of Gemini's edge is simply staying legal. In the Game Arena, a model that couldn't produce a legal
move in four tries lost the game. The paper reports that Gemini models needed fewer than 0.2 retries
per game against most opponents, while Claude Haiku 4.5 and DeepSeek V3.2 often needed more than five.
More in [Can Gemini play chess?](/answers/can-gemini-play-chess/)

## ChatGPT vs DeepSeek and ChatGPT vs Grok

OpenAI's models beat both in the 2025 exhibition. o4-mini, a smaller OpenAI model, beat DeepSeek R1
4–0 in the quarterfinals. In the final, o3 beat Grok 4 4–0, after Grok had looked like the favourite
through the earlier rounds, according to
[Chess.com's report of the final](https://www.chess.com/news/view/kaggle-game-arena-chess-2025-day-3).

The leaderboards tell the same story today. On Dubesor's board, DeepSeek's best model is at 1438 ±99
and Grok's best current reasoning model, Grok 4.20, at 1362 ±89, both below GPT-6 Astra. See
[Grok chess](/answers/can-grok-play-chess/) for Grok's full record.

## Why the "better" chatbot still isn't a good chess player

Whichever assistant you pick, the same limits apply. A language model writes the move that looks most
likely; it doesn't search the replies the way an engine does. It has to keep the board in its head, and
it can lose track in a long game. Even on the best leaderboards, the top models still make illegal
attempts when they only see the list of moves played. We explain the mechanics in
[why ChatGPT plays illegal moves](/answers/why-does-chatgpt-play-illegal-moves-in-chess/), and the
Claude numbers in detail in [What is Claude's chess Elo?](/answers/claude-chess-elo/)

How strong is the top? Dubesor gives a rough, low-confidence comparison from his own tests: about 1850
on his board matched Chess.com's level-16 bots, rated around 2000. So the best chatbot of October 2026
plays like a decent club player at slow time controls, and makes mistakes no engine would.

If you just want to play for fun, both work. Paste the full move list each turn, keep a real board open
to catch illegal moves, and pick the strongest model with thinking turned on. Our guides for
[playing chess with Claude](/guides/play-chess-with-claude/) and
[Can ChatGPT play chess?](/answers/can-chatgpt-play-chess/) walk through it.

## How Chess Coach fits in

Chess Coach doesn't make either assistant a stronger player, and it doesn't play against you. It's a
connector for Claude and ChatGPT that reviews your finished Chess.com and Lichess games with Stockfish
on our own server. Your assistant then explains the key moments in plain words on an interactive board,
with a red arrow for the move you played and a green one for a better move. Every evaluation comes from
the engine, so it doesn't matter which chatbot is the better player. It doesn't work in Gemini or Grok,
and it never helps with games in progress. The free plan includes one in-depth review a day.
[Connect it to Claude or ChatGPT](/connect/), or read the [docs](/docs.html).

## FAQ

### Is Claude or ChatGPT better at chess?

They're close. As of October 2026, OpenAI's GPT-6 Astra is a little ahead of Claude Opus 5.5 on two public LLM chess leaderboards, but their margins of error overlap. Google's Gemini is ahead of both on most boards.

### Is Claude better than Gemini at chess?

Not on current leaderboards. Gemini 3.1 Pro Preview leads Dubesor's AI Chess Leaderboard at 1860 ±85, against 1571 ±97 for Claude Opus 5.5 (October 25, 2026). In August 2025, Gemini 2.5 Pro beat Claude Opus 4 4–0 in Kaggle's AI chess exhibition.

### ChatGPT vs DeepSeek at chess: which is stronger?

OpenAI's models rank above DeepSeek's on the leaderboards we checked. In Kaggle's August 2025 exhibition, o4-mini beat DeepSeek R1 4–0, and on Dubesor's board DeepSeek's best model is at 1438 ±99, below GPT-6 Astra's 1637 ±86.

### ChatGPT vs Grok at chess: who won?

OpenAI's o3 beat Grok 4 4–0 in the final of Kaggle's AI chess exhibition in August 2025. On October 2026 leaderboards, OpenAI's newest models are still clearly ahead of Grok's.

### Can Claude or ChatGPT beat a chess engine?

No. On these leaderboards a chess engine isn't even a rival; it's the referee that judges the AI players' moves. The strongest chatbots play roughly like a decent club player in a slow game, with frequent mistakes.

## Related

- [Can Claude play chess? What it can and can't do (2026)](https://thechesscoach.netlify.app/answers/can-claude-play-chess.md)
- [Can ChatGPT play chess? Elo, tests and limits (2026)](https://thechesscoach.netlify.app/answers/can-chatgpt-play-chess.md)
- [Claude chess Elo: what the LLM leaderboards say (2026)](https://thechesscoach.netlify.app/answers/claude-chess-elo.md)
- [Can Gemini play chess? Gemini's chess Elo in 2026](https://thechesscoach.netlify.app/answers/can-gemini-play-chess.md)
- [Grok chess: can Grok play, and how well? (2026)](https://thechesscoach.netlify.app/answers/can-grok-play-chess.md)
- [Chess MCP server for Claude and ChatGPT: Chess Coach](https://thechesscoach.netlify.app/connect.md)