Ataraxos AI beats Stratego great Pim Niemeijer 15-1 on just 16 GPUs

Machines have been embarrassing human champions for decades. Deep Blue beat Garry Kasparov at chess in 1997, AlphaGo beat Lee Sedol at Go in 2016, and poker bots have been taking professionals’ money for years.

Stratego was the holdout. Now it has fallen too, and the AI that did it was trained for less than the price of a decent used car.

A research team from Carnegie Mellon, MIT, New York University, and Stanford built an AI called Ataraxos. In a 20-game series, it beat Pim Niemeijer, widely regarded as the greatest Stratego player ever, by a score of 15 wins, 1 loss, and 4 draws.

The match and the money

The series took place in July 2025. Ataraxos finished with an 85% win rate against Niemeijer.

In live exhibition matches in August 2025, the system faced attendees at a Stratego championship and went 38 wins to 2 losses, a 95% win rate.

The more surprising number is the cost. Training ran for about one week on 16 Nvidia H100 GPUs. A further four days on four GPUs refined a component called the belief network. The total bill came in under $8,000.

That is a striking contrast with earlier attempts. DeepMind’s DeepNash reportedly required millions in computational costs and still did not reach superhuman play against elite opponents.

Why Stratego was so hard to crack

Chess and Go are games of perfect information. Both players see every piece on the board at all times.

Stratego is different. It involves hidden information, meaning a player cannot see what the opponent is holding. According to the research, the game has more than 10^66 possible configurations.

Ataraxos tackled this with two ingredients. The first is self-play reinforcement learning, where the AI improves by playing against itself millions of times and learning from the results. The second is a generative belief network, which models what the opponent’s hidden pieces are likely to be.

The techniques also carried over to other games. They produced superhuman results in Barrage Stratego, a variant of the game, and state-of-the-art performance in Hanabi, a cooperative card game where players also work with incomplete information.

Background: the long march of game-playing AI

Deep Blue showed raw search power could beat a world champion at chess. AlphaGo showed learned intuition could handle a game too vast for brute force. Poker bots showed machines could bluff and handle uncertainty against human opponents.

The project’s notable contributors include Gabriele Farina of MIT and Samuel Sokota of Carnegie Mellon. An early preprint of the work appeared in November 2025, and a peer-reviewed paper detailing the results was published in Nature on September 30, 2026.

What this means beyond the board

A sub-$8,000 training run that beats a DeepMind effort challenges the assumption that frontier AI milestones require frontier budgets, at least for this class of problem.

There is a hardware angle as well. Ataraxos ran on Nvidia H100s. The project shows those chips can deliver outsized results in small quantities when paired with smarter algorithms.

For human Stratego players, the news is more bittersweet. Niemeijer managed one win out of 20, which is now arguably the most impressive single victory in the game’s history.