Google adopts Werewolf and Poker in AI benchmark 'Game Arena'

Google DeepMind, Google's AI development division, partnered with Kaggle in 2025 to launch Game Arena, a public benchmarking platform where AI models compete in strategy games. On February 2, 2026, it was announced that Werewolf and Poker would be added to the Game Arena's competitive events.

Game Arena: Poker and Werewolf, and Gemini 3 tops chess
https://blog.google/innovation-and-ai/models-and-research/google-deepmind/kaggle-game-arena-updates/

Game Arena was launched in 2025 in partnership with Google DeepMind and Kaggle to measure the reasoning ability and strategic planning of AI models, and until now, chess has been used as the competitive event.

Google releases 'Game Arena,' a benchmark platform for measuring AI game performance - GIGAZINE

Chess is a perfect information game in which all information is public, but in the real world, decisions must be made based on incomplete information. This necessitated the need for a new metric to measure a model's reasoning ability in the face of ambiguity.

The newly added Werewolf game is a team-based social reasoning game that uses natural language dialogue to test how models navigate social dynamics under incomplete information. Google DeepMind says the Werewolf benchmark will help evaluate soft skills, such as communication skills, negotiation skills, and the ability to navigate ambiguity, that are required for next-generation AI assistants.

Werewolf also serves as a safe environment for studying agent safety, allowing us to test the model's ability to detect manipulation by others and its own deception before deploying it in the real world.

Poker, like Werewolf, is a game of imperfect information, but it requires new aspects of risk management and uncertainty quantification from AI models. Models are not only dependent on the luck of the cards they are dealt, but must also predict their opponent's hand and determine the optimal course of action while adapting to their opponent's playing style.

At the time of writing, the latest generation models, Gemini 3 Pro and Gemini 3 Flash, had the highest Elo ratings in both chess and Werewolf. These models demonstrated advanced reasoning capabilities, including identifying inconsistencies between other players' statements and voting patterns and forming consensus.

Google DeepMind is hosting an AI poker tournament in which top models compete under Texas Hold'em rules, with the final rankings scheduled to be published on February 4, 2026.

Google DeepMind also announced that it will analyze the performance of its AI models on these new benchmarks and track their evolution through live-streamed events featuring chess grandmaster Hikaru Nakamura and poker legends.