AI Model That Cost Just $8,000 to Train Beats Stratego World Champion

Ataraxos is a project from researchers at Carnegie Mellon, NYU, Stanford, and MIT. In a series of 20 games, it beat Pim Niemeijer, often called the strongest Stratego player in history, by a score of 15 wins, one loss, and four draws. Training cost less than $8,000. That covered a week on sixteen Nvidia H100 GPUs, plus four more days to train a separate network that estimates which pieces the opponent is hiding. In 2023, DeepMind tried to solve the same problem with a system called DeepNash. That project had a budget of $3 million to $4.5 million and used more than a thousand TPUs. It still lost against the best human players, including that same Niemeijer.

To me, the budget is the real story here. A small team spent a week on sixteen H100s. DeepMind spent several months running more than a thousand TPUs, and still came up short. In my own work with agents, I see this pattern all the time. Quality usually depends on how you frame the task and which model you use, not on how much hardware sits behind it. This has nothing to do with coding directly, but it is an encouraging signal.