Ataraxos AI beats top Stratego pro in low-cost training milestone

A small academic team built a Stratego AI on just 16 GPUs, hinting that elite game agents no longer demand hyperscale budgets.

Researchers from Carnegie Mellon, MIT, NYU and Stanford have built an AI agent, Ataraxos, that decisively beat top Stratego player Pim Niemeijer, winning 15 games, losing one and drawing four. This is notable because Stratego is a long-horizon imperfect‑information game with many hidden pieces, a setting that had resisted prior AI approaches even from DeepMind. The team reports that training Ataraxos required only 16 GPUs and cost just a few thousand dollars, suggesting that state‑of‑the‑art game AIs are becoming accessible without hyperscaler‑level resources. Stratego’s combination of hidden information and extended play makes it a closer proxy than games like chess or Go for many real‑world decision‑making problems where information is revealed gradually.