AI beats top human Stratego players for the first time

Overview

Updated 1 hour ago

An AI called Ataraxos beat the most decorated human Stratego player in history, four-time world champion Pim Niemeijer, by a margin of 15 wins, 1 loss, and 4 draws. It is the first superhuman result in the game's history, published this week in Nature.

Stratego is the hard case for game AI because each player's 40 pieces are hidden. Ataraxos handles that fog of war with a belief network that models the opponent's likely pieces, then refines each move with decision-time search. The whole system trained for about $8,000 — roughly 1/500th the compute of the previous record attempt.

The same design pattern also beat top human players at a Stratego variant and reached state-of-the-art results at two other games. The technique reaches beyond games: business negotiations, military planning, and cybersecurity all involve hidden information.

Why it matters

The technique that beat Stratego's hidden-information fog applies to business, military, and security decisions made with incomplete information.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

85%

Effective win rate in 20-game series

15 wins, 1 loss, 4 draws against four-time world champion Pim Niemeijer.

~$8,000

Training cost

Reinforcement learning on 16 H100 GPUs for a week; belief network on 4 for 4 days.

NVIDIA H100 GPUs used

One week of training, versus a multimillion-dollar compute effort for the prior near-miss.

1/500th

Compute versus prior Stratego effort

1/30th of the self-play games and 1/100th of the training examples.

Voices

Curated perspectives — historical figures and your fellow readers.

Ever wondered what historical figures would say about today's headlines?

People Involved

Gabriele Farina

MIT computer scientist and core Ataraxos developer

Co-developer of Ataraxos

Pim Niemeijer

Four-time Stratego world champion

Lost the 20-game series to Ataraxos, 15-1-4

Max Roelofs

Three-time Stratego world championship runner-up

Commented on the series outcome

Organizations Involved

Massachusetts Institute of Technology (MIT)

Research university

Co-developer of Ataraxos

One of four universities behind Ataraxos, with Carnegie Mellon, New York University, and Stanford.

Nature

Scientific journal

Publisher of the Ataraxos research

Peer-reviewed journal that published the paper describing Ataraxos.

Timeline

2022 October 2026

4 events Latest: 3 days ago

Tap a bar to jump to that date

October 2026
Human Progress reports the result
Latest Report

Human Progress reports the Ataraxos result to a general audience.
3 days ago
September 2026
Nature publishes the Ataraxos paper
Publication

Nature publishes the Ataraxos paper, reporting the first superhuman result in Stratego's history.
5 days ago
2025
Ataraxos defeats Pim Niemeijer 15-1-4
Competition

Ataraxos defeats four-time world champion Pim Niemeijer, 15-1-4, over three weeks.
2025
2022
DeepMind's DeepNash reaches near-top human level at Stratego
Milestone

DeepMind's DeepNash reaches near-top human level at Stratego but cannot beat champions.
2022

Scenarios

Ataraxos pattern goes superhuman in more hidden-information games

Likely Resolves by Q3 2027

Discussed by: Ataraxos's authors, quoted in Nature

The paper presents the belief-network plus decision-time-search pattern as general, and already reports superhuman results at Barrage Stratego and state-of-the-art results at Hanabi and dou dizhu. Applying the same recipe to other imperfect-information games — poker variants, Diplomacy, bridge — is the obvious next step. A follow-up publication would confirm it.

Ataraxos techniques move into real-world decision tools

Possible Resolves by Q1 2028

Discussed by: MIT News and the research team

MIT's announcement names concrete applications: military maneuvers, business negotiations, and cybersecurity. A deployment means a fixed system in a real decision loop, a stiffer test than winning a board game. The trigger is a named organization adopting the belief-network planner.

Rival labs replicate or beat Ataraxos

Uncertain Resolves by End of 2027

Discussed by: Ars Technica and the Nature paper's efficiency claims

Ataraxos's edge — 1/500th the compute, 1/30th the self-play games — invites replication. If an independent group reproduces superhuman Stratego, the result is confirmed; if they beat the margin or the cost, the record moves again. Failure to replicate would leave the claim open.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

May 1997

Deep Blue vs. Garry Kasparov (1997)

IBM's Deep Blue beat world chess champion Garry Kasparov 3.5-2.5. It was the first time a machine won a match against a reigning champion under tournament conditions, at a game with no hidden information.

Then

The victory shocked the chess world and became the defining proof that machines could match human intuition in calculation.

Now

Chess became AI's canonical fully visible game; later champions like AlphaZero showed pure self-play could surpass it without human knowledge.

Why this matters now

Chess's search brute force cannot work in Stratego, where pieces are hidden. Ataraxos needed belief modeling instead of exhaustive lookahead.

2019

Pluribus beats poker pros (2019)

Facebook AI Research and Carnegie Mellon's Pluribus beat top professionals at six-player no-limit Texas hold'em — the first superhuman result in a multiplayer imperfect-information game. It relied on counterfactual regret minimization.

Then

Pluribus beat 13 top pros in a $50,000 prize competition, then won 10,000 hands online.

Now

It showed AI could handle hidden information at poker scale, but poker's state space is far smaller than Stratego's.

Why this matters now

Pluribus proved hidden-information play was possible; Ataraxos extended it to a game with vastly more hidden pieces, using a different mechanism.

2022

DeepNash reaches top-3 at Stratego (2022)

DeepMind's DeepNash reached the top three of online Stratego players using model-free multiagent reinforcement learning, but could not beat champions. The industrial effort cost millions in compute.

Then

DeepNash was a milestone but stopped short of the world's best human players.

Now

It set the benchmark that Ataraxos later crossed, at a fraction of the compute.

Why this matters now

DeepNash is the direct predecessor. Ataraxos is the first Stratego AI to go from near-human to superhuman, using 1/500th the compute and 1/30th the self-play games.