Stratego, the 80-year-old game of hidden pieces and imperfect information, was one of the last competitive board games where humans still won. It is not that anymore.
Ataraxos, an AI built by researchers at Carnegie Mellon, NYU, Stanford, and MIT, has beaten Pim Niemeijer — four-time world champion, holder of 15 Dutch national titles, and the man most qualified humans consider the greatest Stratego player alive — in an official 20-game series, 15 wins to one loss, with four draws.
DeepMind spent an estimated $3–4.5 million and failed to surpass top human players. A university team did it in a week for under $8,000.
What happened
Stratego presented a problem that had resisted AI for longer than most. Unlike chess or Go, both players place their 40 pieces face down, creating more than 10³³ possible setups. A piece's rank is only revealed upon collision. The flag is hidden until someone finds it the hard way.
Poker-derived AI methods had previously handled imperfect information games, but only when the hidden information was modest. Texas Hold'em has 1,326 possible starting hands. Stratego has somewhat more than that.
Ataraxos solved this with a custom GPU simulator and what the authors describe as substantially higher sample efficiency. It trained for one week on 16 Nvidia H100 GPUs, then spent four more days refining its belief network — the part that reasons about what it cannot see, which is most things.
Why the humans care
DeepMind had tried this. Its DeepNash system trained on 1,024 TPU nodes for two to three months — an effort the Ataraxos authors estimate cost $3 to $4.5 million at 2025 prices. It did not beat the best human players. It came close, which in competitive Stratego means it lost.
Ataraxos used roughly 1/500th the compute, 1/30th the self-play games, and 1/100th the training examples. The team was working within academic computing budgets, which is either a lesson in efficiency or a fairly pointed comment on what money has been buying in this industry.
Niemeijer spent more than 600 weeks ranked first in the world. The training run that ended that distinction lasted seven days.
What happens next
The researchers have published in Nature and expressed optimism about the implications for AI reasoning under uncertainty, particularly in domains where information is incomplete and opponents do not announce their intentions.
There are several fields like that. The humans are aware of this. They funded the research.