Chess fell to Deep Blue in 1997. Go fell to AlphaGo in 2016. Poker bots have beaten professionals for years. Stratego, with its hidden armies and long, bluff-heavy games, kept its record — until a team of academic researchers built an AI that beat the best player in history for a few thousand dollars.
The system is called Ataraxos, and its match result, reported in the journal Nature, is lopsided: over 20 online games against Pim Niemeijer — a four-time world champion with more than 600 weeks as the world's top-ranked player — Ataraxos won 15, drew four and lost once. At the 2025 Stratego World Championship, attendees who challenged it fared even worse: the AI won 38 of 40 games.
Stratego's difficulty is structural. Each side hides 40 pieces whose ranks are unknown to the opponent; identities are revealed only in battle. That makes for more than a decillion possible setups and games that can run to 2,000 moves — a vast hidden-information problem, unlike Texas Hold'em's 1,326 possible hands. Earlier AIs, including DeepMind's DeepNash, struggled with the bluffing balance the game demands.
Ataraxos, built by researchers at Carnegie Mellon, MIT, NYU and Stanford, learned by playing itself 163 million times, but added something DeepNash never had: a second "belief" neural network that guesses the opponent's hidden pieces from how they have moved. Instead of grinding through every possible arrangement, the AI samples plausible ones, plays out candidate moves and chooses. Efficiency was engineered in: the team wrote a simulator that runs millions of moves per second on graphics cards.
The cost gap is the headline for anyone outside the board-game world. DeepNash was trained for two to three months on 1,024 specialised Google chips, an effort the authors estimate at $3 million to $4.5 million. Ataraxos needed 16 GPUs for a week plus four GPUs for four days. The researchers say the same architecture also beat three world champions at the faster Barrage Stratego variant, mastered the cooperative card game Hanabi and beat top bots at the Chinese card game dou dizhu — and they are now working on making its strategies explainable.
