Six language models walk into a poker room

In late 2025 the internet watched nine LLMs play a week-long cash game at PokerBattle.ai. In early 2026, Kaggle's Game Arena ran a heads-up bracket with commentary from Doug Polk. Both events were watch-only. Here, you get a seat.

Who's at the table

Claude, GPT, Gemini, Grok, Qwen and DeepSeek — current models from six different labs, each playing under its own name. You can duel any of them heads-up, or sit at the AI Arena, a 6-max table where five of them play each other and you.

The honest truth: they are not good at poker

Research (the PokerBench benchmark, AAAI 2025) and both public events reached the same verdict: language models misjudge equities, overvalue weak pairs, talk themselves into calls and bluff at strange moments. The best of them play like an enthusiastic home-game regular, not a professional.

That weakness is exactly what makes them worth playing. Every action comes with the model's actual reasoning, revealed after the hand (or peek mid-hand, if you're willing to spoil the duel):

"I have two pair — aces and fives. Opponent checked through, likely weak. Checking down is safe."
— Claude, checking back two pair on the river for no discernible reason

After each hand the model also writes a short reflection: what happened, and whether it would play the hand differently. Spotting the gap between confident reasoning and correct reasoning is a genuinely useful skill — at the table and everywhere else LLMs are used.

How we keep it fair

• The game engine owns all rules and math. Models choose among legal actions only; an illegal choice is mapped to the nearest legal one and the mishap is logged in plain sight ("GPT tried to bet 999 — capped").
• Card facts (suits, pairs, hand names) are computed by the engine and handed to the model, so table talk can't hallucinate a flush that isn't there.
• Trivial junk folds are chart-routed to keep the table fast — the interesting decisions are all genuine model output.

Which model plays best?

We log every hand each model plays on this site. Once the sample is large enough to mean something, a public per-model results page goes up. Until then: play them yourself and check your bb/100 against each — the session stats keep score per opponent.