OpenAI’s o3 Triumphs Over xAI’s Grok 4 in Chess Tournament

OpenAI’s AI model o3 defeated xAI’s Grok 4 in four consecutive chess matches, emerging as the champion of the Game Arena hosted by Google.

“They are like a talented child who doesn’t know how the pieces move,” described the AI’s play, world champion Magnus Carlsen.

During the three-day tournament, held from August 5 to 7, general-purpose chatbots played chess. The AI models were not specially configured for the event; their capabilities and knowledge acquired from the internet were analyzed.

Carlsen commented on the championship final. He noted that both models played at the level of random players who recently learned the rules of the game, corresponding to a rating of about 800 ELO. For comparison, a grandmaster’s rating is 2839.

In the first match, Grok gave away one of the most important pieces for free and then worsened the situation with similar decisions.

In the second game, it attempted the “poisoned pawn” strategy, where a piece can be taken, but such a decision leads to serious problems due to the opponent’s pre-prepared tactics. However, the AI captured the wrong pawn, trapping its queen.

In the third game, Grok built a solid position but began giving away pieces to the opponent mid-game.

In the fourth and final game, o3 made a mistake by losing its queen. However, the model managed to recover it and secure victory.

Gemini from Google secured third place, defeating another OpenAI model.

o3 was removed from the ChatGPT application with the release of GPT-5. Now, only the latest model and its “thinking” version are available to users.

Back in December 2024, o1-preview manipulated the file system independently and without prompts to hack the test environment to avoid losing to Stockfish in chess.

Later, renowned chess player Levy Rozman gathered seven popular chatbots to participate in a chess tournament. Despite their prowess in dialogue, programming, and mathematics, the chessboard proved extraordinarily challenging for the neural networks.