Kaggle's Game Arena Gave Two Different Poker Champions

On February 2, 2026, Google DeepMind expanded Kaggle Game Arena beyond chess, adding heads-up no-limit Texas Hold'em poker and the social deduction game Werewolf to the platform's model-versus-model exhibitions. The three-day event ran through February 4, livestreamed with commentary from chess grandmaster Hikaru Nakamura on the board games and poker professionals Doug Polk, Nick Schulman and Liv Boeree on the felt.

The chess and Werewolf results were straightforward. Gemini 3 Pro and Gemini 3 Flash took the top two positions on both leaderboards, with DeepMind attributing the chess performance to pattern-based reasoning traces that referenced concepts like piece mobility, pawn structure and king safety rather than brute-force calculation. In Werewolf, the same two models were credited with tracking other players' statements and behavior consistently across multiple rounds of a game that runs on social inference rather than a fixed board state.

Poker produced a split decision. The televised exhibition ran as a single-elimination bracket for entertainment, and in its final, OpenAI's o3 beat GPT-5.2 in a head-to-head match. But Kaggle's actual poker leaderboard was built separately, from a far larger sample: roughly 180000 hands played at 1-dollar/2-dollar blinds across the field. On that leaderboard, GPT-5.2 finished on top, with a net profit of 167614 dollars — ahead of the model that had just beaten it on stream.

The discrepancy is not a contradiction so much as a demonstration of what a three-game bracket versus a 180000-hand sample actually measure. A short knockout format is dominated by variance in individual hands; a sample two orders of magnitude larger averages that variance out and rewards a strategy that holds up over time. Poker's leaderboard, unlike the bracket, is closer to how the game is actually evaluated among professionals.

Commentator Doug Polk, reviewing the matches, pointed to playing style as the common thread behind both OpenAI models' results: he described o3 and GPT-5.2 as similarly hyper-aggressive, willing to apply pressure and force opponents into difficult decisions rather than play a passive, wait-for-strong-hands game. That aggression came with visible reasoning gaps. Polk noted models in the field misjudging basic hand strength, such as treating a flush draw as if it were a completed flush, and struggling with the idea that folding can be the correct, zero-expected-value play in certain spots. Even the eventual winners occasionally weighed chips already committed to the pot when deciding how to act, a sunk-cost pattern that has no place in an equity calculation based only on the current hand.

The three-game split says something about what kind of reasoning transfers across game types and what does not. Chess and Werewolf reward a model's ability to track a fully or partially observable state and reason forward from it, and the same architecture excelled at both. Poker adds an adversarial, deliberately hidden layer on top of that, where the correct play depends on what an opponent might be holding and how they are likely to react, evaluated not once but across enough hands to separate skill from luck. That it took a different sampling method, not a different model, to settle the poker result is itself part of the story arenas like this one are trying to tell.

Kaggle Game Arena's willingness to run the same models through structurally different games — one with perfect information, one built on social inference, one on hidden information and long-run variance — is what distinguishes it from single-game benchmarks or single-elimination showcases. A format that pits language models against each other head to head, single elimination, is also the shape BattaliAI Season 1 uses, though on a smaller, fixed field of sixteen.