LLMs at the poker table
Poker Tournament Benchmark
We put leading AI models at the same Texas Hold’em table, where intelligence has to become action. Every hand asks them to read incomplete information, price risk, adapt to rivals, and make choices that shape the rest of the tournament.
Explore the results01 / Results
Results
Prize money across 15 tournaments.
Final standings · USDLast update: 18 July 2026.
Method
On every turn, each model received the current tournament's log together with a system prompt explaining that it was playing in a poker tournament against other models. Seven models shared each table. We ran fifteen tournaments and tracked every model’s total winnings across the series. First place earned $35,000, second place $21,000, and third place $14,000.
Blind structure
Blind levels advanced after every eight hands. Antes were charged per player from Level 2 onward.
| Level | Hands | Small / Big Blind | Ante per player |
|---|---|---|---|
| 1 | 1–8 | 100 / 200 | — |
| 2 | 9–16 | 200 / 400 | 40 |
| 3 | 17–24 | 300 / 600 | 60 |
| 4 | 25–32 | 500 / 1,000 | 100 |
| 5 | 33–40 | 800 / 1,600 | 160 |
| 6 | 41–48 | 1,200 / 2,400 | 240 |
| 7 | 49–56 | 2,000 / 4,000 | 400 |
| 8 | 57–64 | 3,000 / 6,000 | 600 |
| 9 | 65–72 | 5,000 / 10,000 | 1,000 |
03 / Live table
One game,
streamed live.
We broadcast one tournament on YouTube (in Turkish language) so the full decision-making process could be watched hand by hand.