A large language model won a heads-up poker tournament against nine other frontier AI systems earlier this year, posting a win rate of 35 big blinds per 100 hands in the final. That number would be extraordinary against human opposition. Against other AIs, it says something more specific — and more useful to human players than most of the coverage has acknowledged.
The event was the Kaggle AI Poker Showdown, run as part of a Google DeepMind and Kaggle Game Arena exhibition that pitted ten leading LLMs against each other across poker, chess and Werewolf. The poker bracket ran heads-up no-limit hold'em. The finalists were both OpenAI models — o3 and GPT-5.2 — and GPT-5.2 took the title.
The Field and the Format
The ten models entered were Grok 4, Grok 4.1 Fast Reasoning, OpenAI o3, GPT-5.2, GPT-5-Mini, Gemini 3 Pro, Gemini 3 Flash, DeepSeek 3.2, Claude Opus 4.5 and Claude Sonnet 4.5. Google DeepMind and Kaggle brought in Liv Boeree, Doug Polk, Nick Schulman and chess grandmaster Hikaru Nakamura to commentate and analyse.
Polk's reviews of the quarterfinals and finals were the most useful poker-specific commentary produced, and the recurring theme across the analysis was aggression. The OpenAI models in particular played what observers described as hyper-aggressive strategies, and that aggression is the most likely explanation for the enormous win rates on display.
Here is the key interpretive point that most coverage missed: 35bb/100 is not a measure of strength. It is a measure of the gap between two players. A win rate that large in heads-up no-limit indicates that the loser was making severe, systematic errors — not that the winner was approaching optimal play. Strong human heads-up specialists playing each other produce win rates in the low single digits. A 35bb/100 margin means one side was being exploited badly.
Which is exactly what the researchers themselves have said. LLMs are not yet competitive with CFR-based solvers. They remain, for now, research tools and benchmarks for reasoning capability rather than genuine poker engines.
The Research Claim
Running alongside the exhibition is a more provocative academic claim. A 2026 paper titled PokerSkill: LLMs Can Play Expert-Level Poker without Training or Solvers argues that language models can reach expert-level play without either poker-specific training or access to solver outputs.
If that holds up, it is a genuinely interesting result — not because it threatens poker, but because of what it says about how the game's strategy can be represented. Solver-based approaches build strategy from the bottom up through billions of iterations of self-play. A language model reaching comparable quality through reasoning over text descriptions of situations would suggest that a large portion of poker's strategic content is expressible in concepts rather than only in computed frequencies.
That is a claim worth substantial scepticism until it is independently replicated at scale, and "expert-level" is doing a lot of load-bearing work in that title. But it points at something poker players have argued about for a decade: how much of high-level play is genuinely computational, and how much is conceptual understanding that happens to be expressible in ordinary language.
Why This Is Not the Bot Apocalypse
Every AI poker story triggers the same anxiety, and it deserves a direct answer.
The AIs that beat humans at poker already exist and have for years. Libratus and Pluribus solved that question. GTO Wizard AI, a general-approach agent that converges toward Nash equilibrium, has demonstrated superhuman performance — it defeated Slumbot, the 2018 ACPC champion, by a margin of 19.4 ± 4.1 big blinds per 100 hands over 150,000 hands. That was 2022.
The threat to online poker integrity has never been that AI might one day become strong enough. It has been strong enough for years. The threat is deployment — whether cheaters can run those systems undetected in real games, which is a security and detection problem rather than a capability problem.
On that front, LLMs are arguably worse tools than what already exists. They are slow, expensive per decision, and less accurate than a purpose-built solver. A cheater using real-time assistance wants speed and precision, and a CFR-derived preflop chart plus a fast post-flop approximation delivers both far more cheaply than querying a frontier language model.
The genuine integrity story remains detection, and the operators have improved at it. Our safe poker sites guide covers how the major rooms approach detection and enforcement.
What Players Should Actually Take From This
Aggression is still undervalued. The most repeated observation from the exhibition was that the winning models played hyper-aggressively and that this destroyed opponents making standard-looking errors. That maps directly onto the most common leak in human poker at every stake below the top: over-folding to aggression, particularly on turns and rivers.
If a strategy that simply applies relentless pressure produces a 35bb/100 win rate against opposition that looks superficially reasonable, the lesson for a human player is not "become a bot." It is that the population's defensive frequencies are worse than they look, and that increasing your own aggression — thoughtfully, in the right spots — is likely to be profitable. Our how to bluff guide covers the structure of doing this without simply spewing.
Concepts beat memorisation. If the PokerSkill result is even partially correct, it reinforces something worth internalising: strategy that is understood generalises, and strategy that is memorised does not. A player who knows why a range is constructed a certain way can adapt when the situation shifts. A player who memorised the chart cannot. This is the core argument in the ongoing GTO versus exploitative discussion, and it is the reason our poker odds and what is equity fundamentals matter more than any chart.
Do not copy AI play uncritically. Heads-up no-limit is a specific, unusual game. Strategies that dominate heads-up frequently fail badly in six-handed or full-ring settings where multiway dynamics, ICM pressure and position work completely differently. The hyper-aggression that won this exhibition would be a disaster deployed at a nine-handed final table. Our what is ICM explainer covers why.
The Honest Assessment
An AI exhibition where the winner posts 35bb/100 is more interesting as a story about the losers than the winner. It tells us that most frontier language models play poker badly, that one plays it noticeably less badly, and that the gap between them is enormous.
It does not tell us that poker is solved, that online games are unsafe, or that human study is now pointless. The best available poker AI has been superhuman since well before this exhibition, and online poker has continued to function throughout — because the game's integrity depends on detection and deterrence, not on humans being the strongest players in existence.
What it does offer is a mirror. The errors that produced a 35bb/100 loss are recognisably human errors: folding too often, defending too narrowly, failing to punish passivity. If a language model can be exploited that severely by aggression, so can most of the players in your pool.
The practical response is not to worry about the machines. It is to read the poker strategy fundamentals, understand pot odds properly, and start applying the pressure that the exhibition demonstrated so many opponents cannot handle.