Ataraxos AI Crushes Stratego Champion Using Just 16 GPUs
Ataraxos crushed Stratego champion Pim Niemeijer 15 to 1 after 163 million self-play games, advancing AI for vast hidden-information challenges at low cost.
Summary
Ataraxos, developed by Carnegie Mellon, MIT, New York University and Stanford researchers, defeated four-time Stratego world champion Pim Niemeijer 15 games to one, with four draws, across 20 online games over three weeks. Niemeijer, ranked world number one for more than 600 weeks, earned $100 per victory and knew the AI would not adapt. Ataraxos also won 38 of 40 challenges at the 2025 Stratego World Championship. Stratego combines 40 hidden piece identities, more than a decillion possible setups, games reaching 2,000 moves and sustained bluffing, compared with 1,326 starting hands in Texas Hold’em and roughly 40 moves in chess.
Ataraxos learned through 163 million self-play games, making large strategy changes early and smaller ones later. A second neural network estimates hidden enemy pieces, allowing the AI to sample plausible layouts and search ahead instead of testing every arrangement. Training cost a few thousand dollars and used 16 GPUs for one week, plus four GPUs for four days. DeepMind’s 2022 DeepNash used 1,024 Google chips for two to three months, an estimated $3 million to $4.5 million at 2025 prices, and played about 34 times more games. A GPU simulator built by Gabriele Farina and lead author Samuel Sokota processed millions of moves per second. The architecture also defeated three Barrage Stratego world champions, mastered Hanabi and beat leading dou dizhu bots. Researchers including Eugene Vinitsky now want explainable strategies before extending the methods toward simplified models of negotiations, financial markets, military conflict and war gaming.
Positives
- Ataraxos beat Pim Niemeijer 15 games to one, with four draws, despite his four world titles and 600-plus weeks atop the rankings.
- Training required 16 GPUs for one week and four more for four days, costing only a few thousand dollars.
- Ataraxos played 163 million self-play games, about 34 times fewer than DeepNash, while producing substantially stronger Stratego results.
- Ataraxos won 38 of 40 public challenges at the 2025 Stratego World Championship.
- The same architecture defeated three Barrage Stratego world champions, mastered Hanabi and surpassed leading dou dizhu bots.
Risks & concerns
- Ataraxos cannot explain why it chooses particular moves, limiting transparency for consequential applications.
- Real negotiations, financial markets and military conflicts lack the fixed rules and clear winners available in board games.
- Stratego’s randomized starting arrangements mean even a theoretically perfect strategy can lose through chance.
- Ataraxos has already shifted Stratego’s metagame, including increased attention to placing flags in corners behind two bombs.