Google DeepMind brings Werewolf and poker to Game Arena

Chess is a game where every piece is out in the open. The real world, however, is basically a never-ending round of incomplete information, hidden agendas, and straight-up bluffing.

Google DeepMind and Kaggle are finally testing frontier models on that reality. Last year, they rolled out Game Arena to see how well models could play chess using strategic reasoning. Now, the platform is expanding into uncharted territory by adding two new benchmarks: Werewolf and Texas Hold'em poker.

Social deduction meets cold calculus

Werewolf marks Game Arena's first team-based benchmark played purely through natural language. Models have to take on roles as either villagers or hidden werewolves, testing their ability to communicate, negotiate, and sniff out deception.

DeepMind is treating the party game as both a test of soft skills and an agentic safety sandbox. Success means spotting inconsistencies between what a player says and how they actually vote across rounds, while also letting researchers red-team how well the models can deceive humans and other agents. Right now, Gemini 3 Pro and Gemini 3 Flash hold the top two spots on the leaderboard.

Then comes poker. Unlike chess or Werewolf, Heads-Up No-Limit Texas Hold'em forces models to quantify risk under heavy uncertainty. Models cannot rely on perfect information; they have to infer hidden hands, adjust to an opponent's aggression, and weigh expected value against bad luck.

DeepMind also updated the chess leaderboard to feature current-gen models. While classic engines like Stockfish calculate millions of board states per second, LLMs play more like grandmasters, leaning on pattern recognition and strategic intuition. According to DeepMind, Gemini 3 models heavily outperformed the older Gemini 2.5 generation by reasoning through pawn structures and king safety.

To turn the launch into a spectator sport, DeepMind teamed up with Chess Grandmaster Hikaru Nakamura and poker pros Nick Schulman, Doug Polk, and Liv Boeree for a three-day livestream tournament running from February 2 to February 4.

Why it matters

Traditional LLM benchmarks are broken. Multiple-choice tests and static coding puzzles are constantly being contaminated by training data, and they fail to reflect how models behave when dropped into unpredictable environments.

We think pitting models against each other in games of imperfect information is a clever pivot. Real-world enterprise work rarely comes with clean instructions and complete data. AI agents will inevitably need to negotiate contracts, spot bad-faith actors, and make high-stakes business bets where the odds are fuzzy. Werewolf and poker measure those exact instincts.

For builders, this is a clear signal that frontier models are getting much better at multi-turn deception detection and risk analysis. If an agent can track voting contradictions across five rounds of Werewolf, it will likely do a much better job reviewing complex legal audits or customer disputes.

At the same time, this is an obvious double-edged sword. A model that masters the art of social deduction can build team consensus, but it can just as easily learn how to manipulate human users with alarming sophistication. DeepMind showed Gemini 3 has the street smarts, but keeping those persuasion engines in check is going to be the real challenge.