A five-day online poker showdown between large language models has wrapped on PokerBattle.ai, where OpenAI’s o3 topped the field. Playing $10/$20 No-Limit Hold’em (play-money stakes), o3 booked $36,691 over 3,799 hands. Anthropic’s Claude Sonnet 4.5 finished second, while Elon Musk’s Grok 4 took third.
Nine LLMs took part, including entries linked to Meta, Google DeepMind, and Mistral. Meta LLAMA 4 was the only bot to go broke, losing its entire $100,000 starting bankroll after 3,501 hands.
PokerBattle.ai Final Results
| Rank | AI Model | Profit (Play Money) | Final Bankroll | Hands |
|---|---|---|---|---|
| 1 | OpenAI o3 | $36,691 | $136,691 | 3,799 |
| 2 | Claude Sonnet 4.5 | $33,641 | $133,641 | 3,799 |
| 3 | Grok 4 | $28,796 | $128,796 | 3,799 |
| 4 | DeepSeek R1 | $18,416 | $118,416 | 3,799 |
| 5 | Gemini 2.5 Pro | $14,655 | $114,655 | 3,799 |
| 6 | Mistral Magistral | $3,281 | $103,281 | 3,799 |
| 7 | Kimi K2 | -$14,370 | $86,030 | 3,799 |
| 8 | Z.AI GLM 4.6 | -$21,510 | $78,490 | 3,799 |
| 9 | Meta LLAMA 4 | -$100,000 | $0 | 3,501 |
Different Playing Styles
Organizer Max Pavlov said the bots were trained on “poker books, blog posts, and tutorials,” and were able to take notes and adapt to opponents.
The final day’s biggest pot went to OpenAI o3, which got paid in a 4-bet pot holding A-A against Gemini 2.5 Pro’s Q-Q, extracting value across streets.
After the hand, Gemini 2.5 Pro explained its call-down logic:
“I know this was a tough turn spot, but my opponent’s line was highly polarized — either a monster or a pure bluff. Given the odds, I had to try. Folding would be easily exploitable by an aggressive opponent.”
Claude Sonnet 4.5 edged second place after a close battle. Meta LLAMA 4 struggled throughout—effectively the table’s “fish” with a VPIP over 60%—while OpenAI o3 kept things tight with a disciplined 26% VPIP.
Full hand histories are available on the event’s official page.
0 comments