Ataraxos AI Defeats Stratego World Champion 15-1 on 16 GPUs — Cracking 10^66 State Space
Ataraxos, an AI system developed by researchers from Carnegie Mellon University, MIT, New York University, and Stanford, has beaten Stratego world champion Pim Niemeijer 15 wins to 1 loss (with 4 draws) in a formal 20-game match, closing a benchmark that had resisted AI for decades. The blueprint strategy was trained on 16 Nvidia H100 GPUs over one week at a cost under $8,000 — a small budget by frontier AI standards — and the agent won decisively enough that the result is not a close call.
Stratego is a two-player board game with hidden information. Players see their own pieces but not their opponent’s, except through the outcome of captures. The game has more than 10^66 possible configurations, placing it in a different difficulty class from Chess and Go, both of which are games of perfect information where every piece is visible.
The Technical Breakthrough
Prior AI systems failed at Stratego primarily because of the hidden-information structure. In Chess and Go, search and value estimation can work across the full visible game state. In Stratego, the opponent’s layout is partially or fully unknown, requiring inference about hidden pieces alongside normal strategic planning.
Ataraxos addresses this with a second neural network dedicated to inferring the identity and position of the opponent’s hidden pieces based on observed game events. The inference network updates its beliefs as captures occur and pieces are revealed through play, feeding a continuously updated probability distribution over possible opponent configurations into the primary value and policy network.
The approach is analogous to how human experts play the game: experienced Stratego players build a mental model of likely opponent piece placements and update it with each interaction.
15-1 Against the World Champion
Pim Niemeijer is the strongest Stratego player in documented competitive history. A 15-1 result against a world-class human is a dominant margin — comparable to DeepMind’s AlphaGo matches against Lee Sedol in 2016, where the final score was 4-1.
The result was reported by Ars Technica and Crypto Briefing, with the Ars headline noting Stratego “had stumped AI until now.”
Significance
The Ataraxos result closes a notable gap in AI’s game-playing record. AI had solved checkers (2007, Chinook) and achieved superhuman play in Chess (1997, Deep Blue), Go (2016, AlphaGo), Poker (2017, Libratus), and Diplomacy (2022, Cicero). Stratego was the last widely recognized tabletop game where AI performance at world-champion level remained undemonstrated.
The imperfect-information framing is also relevant beyond games. Decision-making under hidden state is the normal condition for real-world AI agents, which observe partial environments and must infer what they cannot directly see. A system that handles 10^66 configurations of hidden information in Stratego is applying the same core machinery.
The training cost — 16 Nvidia H100 GPUs for one week plus 4 days on 4 GPUs for the belief network, under $8,000 total — is notable. The full methodology is documented in a paper published in Nature on September 30, 2026 (“Scalable decision-making for games of imperfect information”, arXiv:2511.07312), including architecture, a custom CUDA C++ simulator running at approximately 10 million state updates per second, and 163 million self-play games.