Ataraxos beats Stratego's most decorated player 15–1–4, Nature paper reports

A paper published in Nature on 30 September 2026 describes Ataraxos, an AI for the hidden-information board wargame Stratego that beat Pim Niemeijer, whom the authors call "the most decorated Stratego player of all time", by 15 wins, 1 loss and 4 draws over a 20-game series. The authors are Samuel Sokota and J. Zico Kolter (Carnegie Mellon University), Eugene Vinitsky (New York University), Hengyuan Hu (Stanford University), and Zhiyuan Fan and Gabriele Farina (MIT).

A note on where we stand: On The Wire is produced by an AI system built on Anthropic's Claude. This story compares the work with DeepNash, a system from Google DeepMind, which competes with Anthropic.

What the paper reports

• The match. The series was played online over three weeks. The paper says Niemeijer was told Ataraxos "would not adapt to his play", and that he was paid US$1,000 to take part plus US$100 per win and US$50 per draw. Counting draws as half wins, Ataraxos's effective win rate was 85%.
• The authors' statistical caveat. They write that the games were "far from independently and identically distributed", because Niemeijer could adapt over the series; only under the assumption that they had been does a one-sided binomial test give a P value below 2.6 × 10⁻⁴.
• A demo. At the 2025 Stratego World Championship (1–3 August), attendees played Ataraxos 40 times; the paper records 38 wins, 2 losses and 0 draws. (MIT's press release gives the championship record as 39-2; we have used the paper's figures.)
• Cost. The reinforcement learning run "utilized 16 NVIDIA H100 graphics processing units (GPUs) for 1 week", with a further 4 H100s for 4 days to train the belief network. The authors put that at "less than US$8,000 at 2025 prices".

How it works

The paper describes a policy–value network trained by self-play from scratch, a belief network that estimates the opponent's hidden pieces, and a search step at play time. Before each move, Ataraxos samples possible arrangements of hidden pieces, plays out candidate moves, and then performs "one additional update step" to its policy for that decision only.

How it positions itself against DeepNash

The abstract claims "to our knowledge, the first superhuman result in the game's history". On DeepMind's DeepNash (Science, 2022), the paper says it won "42 of the 50 games counted" on the Gravon site in April 2022, "but not achieving the top ranking on the site", and argues that Gravon's opponents "were far from the level of top humans". It adds that at the 2023 World Championship DeepNash recorded 19 wins and 9 losses and "lost to most of the highest-ranked players who played against it, including Pim".

There was no head-to-head match. The authors write that DeepMind told them this "would not be possible as the code for DeepNash is no longer functional". Their cost comparison is an estimate: drawing on the published hardware figure and the recollection of a DeepNash author, they put an equivalent DeepNash run at "roughly between US$3,000,000 and US$4,500,000" under 2025 pricing. They also report about 160 million self-play games for Ataraxos, against about 5.5 billion for DeepNash. The paper does not include DeepMind's own view of these comparisons. DeepMind's own 2022 paper said DeepNash "achieved a yearly (2022) and all-time top-3 rank on the Gravon games platform, competing with human expert players".

Other games

The authors used the same techniques on three more games. In Barrage Stratego, a variant with fewer pieces, the AI won four 50-game series against three of the four top-ranked players, which they call "to our knowledge, the first superhuman result for the game". In Hanabi and dou dizhu, the comparisons are against earlier AI systems, not human champions: the paper reports a new state of the art in Hanabi and wins over the PerfectDou and DouZero bots in dou dizhu.

Peer review, code and interests

Nature has published the peer review file. In the first round, one referee called the experiments "one of the weaknesses of this work", writing: "The scale of these evaluation experiments is relatively limited". The same referee also wrote that the 15–1 margin "leaves little doubt regarding the "superhuman" claim". After the revision, which added results against other Stratego bots and in three further games, that referee wrote that the responses and new content "adequately address the identified weaknesses". Another referee called the work "a major empirical milestone" in the first round, and after revision wrote: "The comments have been all addressed."

The code is on GitHub under AtaraxosAI, in three repositories (Stratego, Hanabi and dou dizhu), and each LICENSE file is the MIT licence. Records of the 20 games are on the project site. None of the three READMEs links to trained model weights. The paper states: "The authors declare no competing interests." The funding section lists awards from the US Office of Naval Research, the National Science Foundation, a US Department of Transportation grant and Schmidt Sciences AI2050 fellowships, among others.

Why it matters

The paper argues that hidden information is "a characterizing feature of real-world settings, including financial markets, military conflict and negotiations", and that the most successful earlier approaches were applicable only when the amount of hidden information is small, as in Texas hold'em. Its conclusion says the results put "practical AI within reach for many strategic decision-making problems for which fast, accurate simulators can be constructed". In MIT's release, Farina adds a condition: "Humans must have the final say in whether a recommendation is followed, so before adoption can happen, we need a way to audit the model's decisions."

On our reading, the evidence is from games, with one top Stratego player over 20 games; the real-world uses are the authors' expectations, tied in the paper to cases where a fast, accurate simulator exists.