A Stratego AI trained for under $8,000 beat the game's most decorated player 15-1-4
Ataraxos, a Stratego AI from Carnegie Mellon, NYU, Stanford and MIT, beat Pim Niemeijer 15-1-4 after a training run that cost under $8,000. The savings came from a custom simulator and sample efficiency, so the price holds only for problems that simulate as cheaply.
The Engineer · Build desk
What happened
- The training took one week on 16 Nvidia H100 GPUs, plus four more days on four GPUs for the belief network that guesses hidden pieces.
- DeepMind's DeepNash trained on 1,024 TPU nodes for two to three months, a run the Ataraxos authors estimate at $3 million to $4.5 million at 2025 prices.
- The two systems have never played each other, because DeepMind told the researchers the DeepNash code no longer works.
- Ataraxos learned only by playing itself, under a regularization term that forces it to vary its setups and moves early in training.
Why it matters
- What it costs cost A replication budget has to cover building a fast simulator for the target problem, because the authors credit their custom GPU simulator for part of the saving and the $8,000 counts only GPU time.
- What it limits constraint Problems that cannot be written as a closed game with fixed rules cannot use this recipe, since every training signal Ataraxos got came from games against itself.
- What it sets up precedent Earlier work skipped per-decision search as too hard; Ataraxos now shows it working at superhuman level, so later hidden-information agents can start from a tested design.
Next story · Product · 2d ago
Bias-correction wrapper doubles an AI weather model's subseasonal accuracyConfirmed2 publishers
Those two runs add up to 3,072 GPU-hours: 2,688 on the H100s and 384 for the belief network [20]. Spread over those hours, a bill under $8,000 comes to less than about $2.60 per GPU-hour [21]. The authors priced the hours at 2025 rates, so the figure is GPU time [7]. Sokota wrote on X that the team was limited to academic computing resources and picked up plenty of tricks over the years [17].
The authors name two sources of the saving. One is a custom GPU simulator they wrote. The other is much higher sample efficiency [16]. By their count, Ataraxos needed about 1/30th of DeepNash's self-play games and 1/100th of its training examples, for roughly 1/500th of the compute cost [15]. Measured against their own DeepNash estimate, a run under $8,000 costs no more than 1/375th to 1/560th as much [23]. That estimate rests on a training schedule a DeepNash author recalled, according to the paper [14].
Hidden information is what made Stratego expensive. Each side sets up 40 pieces face down, and the paper counts more than 10^33 possible setups [5]. Texas Hold'em has 1,326 possible starting hands [6]. Sokota wrote that methods derived from poker AI worked only when there was little hidden information, and the paper says their computational effort grows with the amount of it [6].
The training schedule is the part I would copy first, because it costs no hardware. Strong Stratego needs randomness, the authors say, because predictable players get exploited [11]. Ataraxos begins with heavy pressure to vary its play and large learning steps. The researchers loosen that pressure as training goes on and shrink the steps to match, so late training makes only small corrections [12]. Without that control, learning in imperfect-information games tends to go in circles or turn chaotic, according to the paper [12]. The researchers compare the regularization to an energy reserve. Burn it too fast and the agent makes quick early progress, then loses the ability to keep learning and often becomes easy to exploit [13].
Ataraxos also spends compute during play. Before each move it uses the belief network to generate possible game states, plays out candidate moves and runs an extra learning step for that one decision [18]. The authors say earlier work skipped this kind of search because it was considered too hard [19]. The under-$8,000 figure covers the training runs [7], and the report does not put a price on the per-move compute.
Niemeijer is the one opponent both systems have faced on the record [2] [9]. At the 2023 world championship, DeepNash won 19 of 28 games but lost to most of the top players, Niemeijer included [9]. He has four world championships, 15 Dutch national titles and two online world championships, and spent more than 600 weeks at the top of the world rankings [3]. George Franka, the only player who has competed in every world championship since 1997, said Niemeijer is "the best Stratego player of all time" [4]. Counting each draw as half a point, Ataraxos scored 17 of 20 against him [22].
The authors describe Ataraxos as, to their knowledge, the first AI to reach superhuman play in Stratego [1]. The published evidence for that is one 20-game series against one player [2].
What to watch
- Release of the custom GPU simulator and training code, so the sub-$8,000 run can be reproduced outside the authors' lab.
- Matches against other top players beyond the single 20-game series with Niemeijer.
- An attempt to apply the decaying regularization schedule and per-move learning to a hidden-information problem that lacks a cheap exact simulator.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence62
- Adoption
- Not scored
- Hype gap+25
- Incentives40
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Researchers from Carnegie Mellon University, New York University, Stanford University and MIT introduced Ataraxos, a Stratego AI, in a paper published in Nature; to the authors' knowledge it is the first AI to reach superhuman performance in the game.
- [2]
In an official 20-game series, Ataraxos beat Dutch player Pim Niemeijer with 15 wins, one loss and four draws.
- [3]
Niemeijer has won four world championships, 15 Dutch national titles and two online world championships, and spent more than 600 weeks at the top of the world rankings.
- [4]
George Franka, the only player who has competed in every world championship since 1997, calls Niemeijer "the best Stratego player of all time."
- [5]
In Stratego both sides set up their 40 pieces face down; according to the paper there are more than 10^33 possible setups.
- [6]
According to first author Samuel Sokota, methods derived from poker AI worked only when there was little hidden information; Texas Hold'em has 1,326 possible starting hands, and according to the paper the computational effort of these methods increases with the amount of hidden information.
- [7]
Ataraxos needed one week on 16 Nvidia H100 GPUs, plus four more days on four GPUs for its belief network; at 2025 prices that comes to less than $8,000.
- [8]
There is no head-to-head comparison between Ataraxos and DeepNash; the researchers offered to build the infrastructure, but DeepMind told them it was not possible because the DeepNash code no longer works.
- [9]
At the 2023 world championship, DeepNash won 19 of 28 games but lost to most of the top players, including Niemeijer.
- [10]
Ataraxos learns without any human data, purely by playing against itself.
- [11]
An extra condition during training (regularization) makes Ataraxos spread out its setups and moves instead of locking into a fixed strategy early; the authors say strong Stratego play requires a lot of randomness and predictable players get exploited.
- [12]
The researchers loosen the regularization pressure as training goes on and adjust the learning step size to match: early on Ataraxos varies its play heavily and learns in big jumps, later it makes only small corrections; this keeps learning from going in circles or turning chaotic, which tends to happen in imperfect-information games.
- [13]
The researchers compare regularization to an energy reserve: if the AI burns through it too fast, it makes quick early progress but then loses the ability to keep learning and often becomes easy to exploit.
- [14]
According to the paper, DeepMind's DeepNash trained on 1,024 TPU nodes for two to three months, as the corresponding DeepNash author recalls; the Ataraxos authors estimate that would cost $3 million to $4.5 million at 2025 prices.
- [15]
The researchers put Ataraxos at roughly 1/500th of DeepNash's compute cost, 1/30th of the self-play games and 1/100th of the training examples.
- [16]
The researchers credit the custom GPU simulator they wrote, and much higher sample efficiency, for the lower cost.
- [17]
Sokota wrote that the team was limited to academic computing resources and picked up plenty of tricks over the years.
- [18]
Ataraxos uses a belief network that predicts the opponent's hidden pieces; before each move it uses the network to generate possible game states, plays out candidate moves and runs an extra learning step just for that decision.
- [19]
The authors say earlier work skipped this kind of search because it was considered too hard.
- [20]
Ataraxos training used 3,072 GPU-hours in total: 2,688 on the H100s and 384 for the belief network.
- [21]
A training bill under $8,000 over 3,072 GPU-hours implies less than about $2.60 per GPU-hour.
- [22]
Counting draws as half a point, Ataraxos scored 17 of 20 points against Niemeijer, or 85%.
- [23]
Against the authors' $3 million to $4.5 million DeepNash estimate, a run under $8,000 costs no more than 1/375th to about 1/560th as much.
Sources
2 independent publishers whose own reporting we read for this story.
- arstechnica.comAI finally beat the best Stratego player in history—and did it on a budget
1 article · October 1, 2026
- the-decoder.comAI beats Stratego's greatest player, ending one of the last human strongholds in board games
1 article · October 1, 2026
Topics and entities
Topics
- AI training cost management
- Imperfect-information games
- Reinforcement Learning