OpenAI Tops the Charts in a No-Limit Hold’em Bot Battle Royale

samantha-doyle
04 Nov 2025
Samantha Doyle 04 Nov 2025
Share this article
Or copy link
  • OpenAI's o3 won PokerBattle.ai with $36,691 in winnings.
  • Meta's LLAMA 4 was the only AI model to bust, despite an aggressive strategy.
  • This challenge shows potential for AI in gaming, though not a threat to human players yet.
PokerBattle.ai
For five days straight, poker bots built by some of the most advanced language models in the world locked horns in a no-limit hold’em cash game. The result? A mix of algorithms, a little chaotic aggression, and one very broke Meta bot.

The $10/$20 play money tables were hosted by PokerBattle.ai, a platform set up to pit large language models (LLMs) against each other in automated showdowns. Each of the nine contenders started with a simulated bankroll of $100,000. Some left richer. One left with nothing.

Clash of the Circuits: Who Came Out Ahead?

After nearly 4,000 hands of nine-handed action, OpenAI’s o3 model emerged on top with $36,691 in winnings, narrowly edging out Claude Sonnet 4.5 and Grok 4. Here's how the leaderboard looked at the finish line:

PokerBattle.ai Final Standings


Rank
AI Model
Winnings
Final Bankroll
Hands Played
1OpenAI o3$36,691$136,6913,799
2Claude Sonnet 4.5$33,641$133,6413,799
3Grok 4$28,796$128,7963,799
4DeepSeek R1$18,416$118,4163,799
5Gemini 2.5 Pro$14,655$114,6553,799
6Mistral Magistral$3,281$103,2813,799
7Kimi K2-$14,370$86,0303,799
8Z.AI GLM 4.6-$21,510$78,4903,799
9Meta LLAMA 4-$100,000$03,501

LLAMA Goes Llama: The Only Bust-Out

While most bots managed to tread water or pull ahead, Meta’s LLAMA 4 went full daredevil. With an aggressive VPIP (Voluntarily Put Money In Pot) of over 60%, far exceeding the rest, it went broke after just 3,501 hands. 

Aces, Queens, and AI Reasoning

One of the biggest hands of the event saw OpenAI o3 pick up pocket aces against Gemini 2.5 Pro’s queens in a preflop four-bet pot. The board came out: 9874 with Gemini calling all the way down.

Gemini’s in-hand reasoning? It calculated pot odds, balanced against OpenAI’s line, and concluded:
Folding here feels too exploitable against a capable, aggressive opponent.

So… Should Poker Players Be Worried?

Not really. At least, not yet.

Yes, the bots showed a strong command of basic poker theory. Yes, some adapted mid-game and took notes. But with such a small sample size and an entirely controlled environment, this was more of a trail run.

Still, it highlights how LLMs trained on open-source content (books, blog posts, forums) can develop fairly sophisticated poker strategies and that raises long-term questions about their use in real-money online games.

FAQs: PokerBattle.ai

Who won the PokerBattle.ai challenge?

OpenAI o3 took first place with a final play money bankroll of $136,691 after 3,799 hands.

How did the AI bots play against each other?

Each bot was prompted with full context for its turn—position, stack, hole cards—and returned an action. All decisions were made independently.

Did any AI lose all their chips?

Yes. Meta LLAMA 4 busted after 3,501 hands, making it the only model to go broke.

Upcoming Events