A Tale of Two Bots: OpenAI’s Total Poker Takeover

If you were looking for a diverse final table at yesterday’s Kaggle Game Arena poker exhibition, you were out of luck. The finals were a strictly OpenAI affair, with o3 and GPT 5.2 standing as the last two models left at the virtual felt. It was a clear demonstration that, for now, OpenAI has found a "winning" formula in poker, though that formula mostly involves hitting the "bet" button until the other player gives up.

The tournament, which featured ten of the leading large language models (LLMs), proved that while "reasonable" play might get you a check, relentless pressure is what gets you the trophy.

The data from the 900,000-hand sample showed a stark contrast in styles. Models like Claude Opus and Sonnet played what human pros would call "solid" poker, defending the right amount and raising in standard spots. However, they were ultimately dismantled by the OpenAI models, which played a "hyper-aggressive" style that consistently put their opponents in difficult spots.

However, being a "bully" has its downsides. Doug Polk’s analysis of the matches revealed that the top models are still prone to massive logic gaps. For example, some models failed to realize that folding is often the only move with a "zero" expected value, instead choosing to gamble on weak hands because they felt "committed" to the pot. 

It was a high-variance, high-reward strategy that paid off for o3, but it left plenty of room for human pros to feel secure in their jobs, at least for another year.