Top AI models from OpenAI, Google, and Anthropic to compete in public chess tournament

[Skye Jacobs]

In a nutshell: Several of the world's most advanced artificial intelligence models will soon face off in a public chess tournament. The three-day event will feature leading AIs from OpenAI, Google, Anthropic, xAI, DeepSeek, and Moonshot as they compete for chess supremacy in a high-profile exhibition.

The inaugural tournament marks the launch of Kaggle's Game Arena, a new platform designed to test and compare the decision-making and reasoning skills of large language models across a variety of strategic games. Organized in collaboration with Google DeepMind, Chess.com, and chess streamers Hikaru Nakamura and Levy Rozman, the event will stream both matches and expert commentary.

Eight AI models, including OpenAI's o3 and 04-mini, Google's Gemini 2.5 Pro and Gemini 2.5 Flash, Anthropic's Claude Opus 4, xAI's Grok 4, DeepSeek-R1, and Moonshot's Kimi 2-K2-Instruct, will compete in a single-elimination bracket. Each matchup will use a best-of-four format, beginning with quarterfinals and culminating in a championship round.

Every move made by the AIs will be the result of independent reasoning: models will not be allowed to use external chess engines or resources, and their interaction will be limited to text only. They will not receive a list of legal moves and must instead generate responses entirely on their own, with a strict 60-minute time limit per move. Persistently illegal moves may result in automatic forfeiture.

Beyond the public tournament, Kaggle plans to build a comprehensive leaderboard using a much larger set of simulated matches conducted behind the scenes. This expanded ranking system will assess model performance across hundreds of additional games, with randomly assigned matchups to enable broad and rigorous comparison.

Chess was selected for the initial showcase because, according to Google, the game remains resistant to being "solved" outright. Each match presents unique challenges that require players – human or artificial – to adapt, strategize, and anticipate their opponent's plans.

While the first Game Arena event focuses on chess, future competitions are expected to feature games such as Go and Werewolf. These titles were chosen not only for their complexity, but also for their ability to simulate real-world skills such as long-term planning, reasoning under uncertainty, and collaborative decision-making.

The matches are likely to attract interest from gaming enthusiasts and technology observers alike, offering insight into how these systems approach strategy, learning, and competition. Over time, the initiative aims to incorporate a wider variety of strategic, cooperative, and real-world simulation challenges to further test and expand the capabilities of AI models.

Permalink to story:

Top AI models from OpenAI, Google, and Anthropic to compete in public chess tournament

 

[godrilla]

Instead of competing with Humans ai should be teaching people how to play chess and help those who are advanced players on becoming grand masters imo.

 

[brucek]

Are these models even that good at chess? And how does "60 minutes" compare to the time allowed for human player moves?

I'm not an AI expert but this feels like the wrong tool for the job. And in my limited attempts at playing games with/alongside of ChatGPT, while I've seen it demonstrate familiarity and some good ideas with for example Slay the Spire, it also frequently gets basic rules wrong and can repeat the same type of mistake over and over despite being corrected each time.

 

[mbk34]

I'm not an AI expert but this feels like the wrong tool for the job.

It's completely the wrong tool for the job. I can't see what this competition would prove. Perhaps it simply proves that the organisers don't understand the concept of general AI. I suspect all it will prove is that AI doesn't implicitly know how to play chess. I would be impressed though if one of the AI's simply cheats and creates an interface to AlphaZero (or a chess database) to get the best moves.

 

[stewi0001]

Are these models even that good at chess? And how does "60 minutes" compare to the time allowed for human player moves?

I'm not an AI expert but this feels like the wrong tool for the job. And in my limited attempts at playing games with/alongside of ChatGPT, while I've seen it demonstrate familiarity and some good ideas with for example Slay the Spire, it also frequently gets basic rules wrong and can repeat the same type of mistake over and over despite being corrected each time.

I think it is more to see how these AI models handle a random situation. Like one fun thing to do is ask them how may times the letter R appears in Strawberry. AIs do not have a hardcoded memory. It is more about patterns and predictions based on said patterns. In a sense, LLM AIs have alzheimer's if you do not have a cached conversation.

 

[jeffbert]

Queen to Queen's level 3.

 

[jeffbert]

So, they must give the AI both the rules & the object of Chess, & then see which best understands them. Forget playing one against the other, until the get that right.

I was a part-time Chess instructor teaching at a food court & in Elementary schools before or after school. It took a long time to get the kids to play without making illegal moves. It may seem an insult to AI to compare it to 10 year-old kids, but tough!