With most information hidden, the game Stratego had stumped AI—until now

A team of researchers from Carnegie Mellon, MIT, New York University, and Stanford University has built an AI called Ataraxos that beat Pim Niemeijer, described as arguably the best Stratego player of all time, 15 games to one, with four draws. The team says training took just 16 GPUs for a week, plus four additional GPUs for four days to train a belief model, at a cost of a few thousand dollars.

Stratego is an imperfect-information game in which each player gets 40 pieces representing military ranks, from a marshal down to a spy, plus bombs and a flag. Players see where opponent pieces are but not what they are; identities are revealed only when pieces collide in battle. That hidden information can exceed a decillion possible setups, and games can easily last 2,000 moves, compared with roughly 40 moves in chess.

Ataraxos learned through self-play, playing 163 million games. The team says it adjusted strategy with large changes early in training and small ones later. A second neural network, a belief model, guesses the opponent's hidden pieces from their movement, allowing Ataraxos to sample plausible arrangements rather than iterate through every possibility. The researchers say DeepMind's DeepNash, introduced in 2022, lacked such a pre-move search.

Over three weeks, Niemeijer played 20 online games against Ataraxos, earning $100 for each win and winning once. Researchers attributed that loss to luck, noting Stratego requires randomizing piece arrangements, and said Ataraxos also benefited from luck. At the 2025 Stratego World Championship, attendees who challenged Ataraxos fared worse, with the AI winning 38 of 40 games.

DeepNash was trained for two to three months on 1,024 of Google's specialized chips, which the Ataraxos team estimates would cost $3 million to $4.5 million at 2025 prices. The same architecture also beat three world champions at Barrage Stratego, mastered the cooperative card game Hanabi, and beat the best bots at dou dizhu. The work is published in Nature, 2026, DOI: 10.1038/s41586-026-11036-y.