OpenAI might’ve deprecated o3 during its GPT-5 launch yesterday, but the model has made the company proud on its way out.
OpenAI’s o3 model has won Kaggle’s AI chess tournament, which saw AI models from companies like OpenAI, Google and Anthropic and xAI take part. In the final, o3 beat Grok 4 four games to nil. Gemini Pro 2.5 beat o4-mini to win the bronze.
Now computers have been better than humans at chess for a while now, but these programs, such as Stockfish, are specially designed to play chess. On the other hand, this Kaggle competition pitted general purpose AI models against each other.
As it turned out, these models weren’t particularly good at chess. The models frequently blundered, or hung pieces for opponents to take. Some weren’t able to play legal moves at all, and were disqualified for making four consecutive illegal moves. But given how these models weren’t trained to play chess at all, what they managed was still impressive.
The first rounds saw 4-0 results in all matches. o3 easily beat Kimi K2, which was unable to play very much. Kimi K2 was the only non-thinking model in the line-up, so its below-par performance was expected. In the other first-round matches, o4-mini beat DeepSeek R1, Gemini 2.5 Pro beat Claude Opus 4, and Grok 4 beat Gemini 2.5 Pro Flash.
The first semi final was again one-sided, with o3 beating o4-mini 4-0. But the second semi-final provided to be a thriller. Gemini 2.5 Pro won the first game against Grok 4, but blundered in two subsequent games to lose them both. Grok 4 then blundered in the 4th game, taking the match to a tie break. Gemini seemed to have the upper hand for the longest time, until it hung its queen, and Grok 4 was quick to capitalize. The match lasted more than 50 moves, and was perhaps the most interesting of the tournament.
The final saw o3 take on Grok 4, but the result was one-sided, with o3 winning all 4 matches. OpenAI’s o3 stamped its authority on the tournament by winning the title undefeated.
The matches were livestreamed on YouTube, and were commented on by top chess Grand Masters like Hikaru Nakamura and Magnus Carlsen. What made the matches even more interesting is that viewers could read the reasoning traces of some models like Gemini 2.5 Pro, which showed exactly what the model was thinking, which moves it was considering, and why it chose a particular move. This is something that viewers can’t see when humans play chess, and can see only the final moves played by the players. As these models get better, it’s likely that such AI tournaments could become more popular, especially with a clutch of companies all vying amongst themselves to create the best AI models. And while OpenAI has stamped its authority this time, xAI and Google will be itching to get their own back in subsequent tournaments.