Credits pay for questions to the AI coach. They never expire.
The AI Chess Arena lets you do something that was science fiction a few years ago: sit down at a chessboard opposite a large language model — the same kind of AI that powers chatbots — and see how it actually plays. You can challenge a model yourself, or assign an AI to both colors and watch two of the world's most advanced language models battle each other move by move.
The current roster includes OpenAI's ChatGPT 5.6 SOL, Grok 4.5 from xAI, Anthropic's Claude Sonnet 5, Claude Opus 5.0 and Claude Fable 5 Max, DeepSeek 4.0, and Google's Gemini Pro 3.1, Gemini Flash 3.5 and Gemini Deep Think. New frontier models are added to the arena as they are released.
Unlike typical experiments where an LLM is asked to predict the next move via a chat prompt, this arena takes a completely different approach. As detailed in the About button, each language model was given a single programming task: to write a complete, self-contained JavaScript chess engine from scratch. They were allowed to use the chess.js library to handle board state and rule validation, but they had to design their own evaluation functions and search algorithms (like Minimax or Negamax).
What you are watching are actually custom chess engines built entirely by these language models in a single response. The models aren't guessing the next move live over an API; rather, the JavaScript code they generated runs locally in your browser's Web Workers, calculating moves up to the search depth you specify.
Traditional engines like Stockfish are the result of decades of human fine-tuning and testing. When an LLM is asked to write an engine from scratch in one shot, the result is fascinating but imperfect. While most frontier models successfully implement standard algorithms like alpha-beta pruning, their custom evaluation functions—how they value piece placement, pawn structure, and king safety—vary wildly.
Some models write highly aggressive engines, while others prioritize solid structures. They might miss a tactical sequence because their code isn't optimized enough to search deeply in a reasonable time, or because their handwritten evaluation logic has blind spots. Pitting these AI-generated engines against each other is a surprisingly revealing benchmark of how well different models understand game theory, algorithms, and software architecture.
Use the two player selectors to choose who controls each color. Set one side to Human to play against the model of your choice, or set both sides to AI models, start the game, and spectate. The panel tracks move history and captured material for both sides, and you can flip the board, undo moves in a human game, or reset and try a different matchup at any time. If you want to know which model is currently the strongest, the fastest way to find out is to stage a head-to-head and watch it unfold.
When you are done experimenting with the machines, test yourself: play a real game against the Stockfish engine at your own level, get an estimate of your ELO rating, or sharpen your tactics with our chess puzzle trainer.