Researchers from Carnegie Mellon, MIT, New York University and Stanford say their AI, Ataraxos, beat Stratego champion Pim Niemeijer in a 20-game online match. The work is published in Nature. Stratego hides the identity of each player’s pieces, which made it one of the last classic board games where top humans still beat machines.
New to this? Read it in simple words
- University researchers built an AI that plays the board game Stratego.
- It beat the world’s top player 15 games to one, with four draws.
- In Stratego you cannot see what the other player’s pieces are.
- The team says training cost only a few thousand dollars.
- Self-play
- An AI learns a game by playing many games against itself.
- Hidden information
- Facts in a game that one player cannot see, like cards in poker.
- GPU
- A graphics chip; AI models are trained and run on them.
The match
Niemeijer played 20 online games against Ataraxos over three weeks and won once, Ars Technica reports. The AI won 15, and four were drawn.
He has four world titles and more than 600 weeks as the top-ranked player, according to the researchers. He earned $100 for each win.
At the 2025 Stratego World Championship, Ataraxos won 38 of 40 games against players who challenged it, Ars Technica reports.
From the Nature paper as reported by Ars Technica, and the arXiv version.
How it works
Ataraxos learned by playing 163 million games against itself. A second neural network guesses the opponent’s hidden pieces from how they move.
That lets the AI look ahead before each move by sampling likely positions, something DeepMind’s 2022 system, DeepNash, did not do.
“For machines, it’s easy,” said MIT’s Gabriele Farina, a co-author, about ignoring a secret the machine knows when deciding a move.
The budget
Training used 16 GPUs for a week, plus four GPUs for four days for the guessing model, the team says. Ars Technica puts the cost at a few thousand dollars.
The team estimates DeepNash’s training would cost $3 million to $4.5 million at 2025 prices. That is the team’s comparison, not DeepMind’s.
A first version of the paper appeared on arXiv in November 2025. The Nature paper is the peer-reviewed version.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1
- 2Research · October 1, 2026With most information hidden, the game Stratego had stumped AI—until now Ars Technica
- 3Research · October 2, 2026This AI Finally Beat the Best Humans at One of the Last Board Games We Still Ruled ZME Science
- 4Primary source · November 10, 2025Superhuman AI for Stratego Using Self-Play Reinforcement Learning and Test-Time Search arXiv
We read Ars Technica in full, compared ZME Science, and checked the match result and hardware against the arXiv version of the paper. Nature’s page did not open for us, so we cite it by its DOI. Outlets give different ratios for how many fewer games Ataraxos needed than DeepNash, so we leave the ratio out. The cost comparison is the team’s estimate.