GPT-o3 played chess with Grok: who won?

AI developers have organized the first-ever three-day chess competition between their large language models. The Kaggle AI Exhibition Tournament brought together 8 representatives of Anthropic, DeepSeek, Google, and Moonshot AI OpenAI and xAI. The chess tournament between AIs took place on August 5-7.

The models of Chinese developers DeepSeek and Moonshot AI were eliminated at the quarterfinals stage. The third place in the tournament went to the model Google Gemini who won GPT o4-mini with a score of 3.5:0.5. Representatives of OpenAI and Elon Musk’s xAI faced off in the final.

Interestingly, initially, the favorite was considered to be new xAI Grok 4 model.

“Until the semifinals, it seemed that nothing could stop Grok 4 from winning,” said Chess.com columnist Pedro Piñata.

Encouraged by these successes, even before the final, Elon Musk wrote in X (formerly Twitter):

“By the way, this is a side effect. @xAI spent almost no effort on chess.”

As a result, in the final, the OpenAI GPT o3 model won a confident 4:0 victory over the xAI Grok 4.

“On the last day, the illusion was dispelled. The talkative o3 simply dismantled his mysterious opponent, scoring four convincing victories. Grok’s game became unrecognizable — mistakes occurred quickly and often. And o3 didn’t leave a single chance,” Pedro Piñata added.

In the final, Grok 4 made a series of mistakes, including losing the queen several times. The OpenAI model, on the other hand, demonstrated a stable level of play. The average correct move rate in 12 games was 90.8%. In Grok 4, it was 80.2%.

It is worth recalling that Sam Altman and Elon Musk once co-founded OpenAI, but Musk later left the company to develop his own xAI laboratory.

Developers use such tournaments as a testing ground for algorithms in tasks that require logical thinking, strategic planning, and working with strict rules. Chess, like other complex board games, allows to check how efficiently artificial intelligence can learn, predict the opponent’s actions, and make optimal decisions.

Source: cybernews