Academic AI Crushes Stratego World Champion For Under $8,000

A joint team from Carnegie Mellon, NYU, Stanford, and MIT developed an AI named Ataraxos that achieved superhuman performance in Stratego, a board game containing over 10^33 possible setups. The model defeated Pim Niemeijer, a four-time world champion considered the game's greatest player, with an unprecedented 15 wins, one loss, and four draws in a 20-game official series. Training the AI cost under $8,000 using 16 Nvidia H100 GPUs over a single week, plus four days for a belief network, operating at roughly 1/500th the compute cost of prior failed attempts. DeepMind’s DeepNash previously failed to defeat top human players despite training on 1,024 TPU nodes for up to three months—a process estimated to cost between $3 million and $4.5 million today. By forcing the AI to spread out its moves and heavily varying its play early in training, the researchers prevented the system from becoming predictable in an environment strictly defined by imperfect, hidden information.