Poker, Meet the Algorithm: OpenAI o3 Dominates the LLM Cash Game

samantha-doyle
04 Nov 2025
Samantha Doyle 04 Nov 2025
Share this article
Or copy link
  • OpenAI o3 wins with $36,691 profit; Meta LLAMA 4 goes broke.
  • AI models show human-like behavior: errors, overconfidence.
  • Raises questions on AI in real-time poker assistance.
PokerBattle.ai
What happens when you drop nine AI models into a poker room and tell them to battle for five days straight? You get a surprisingly human outcome: overconfidence, mistakes, and a few flashes of brilliance.

The PokerBattle.ai experiment saw leading large language models (LLMs), including OpenAI o3, Claude Sonnet 4.5, Grok 4, and Meta LLAMA 4, grind through nearly 3,800 hands of $10/$20 no-limit hold’em, each starting with a $100,000 virtual bankroll.

Winners, Losers, and Digital Tilt

After the cards settled, OpenAI’s o3 came out ahead with a $36,691 profit, narrowly followed by Claude Sonnet and Grok. Only one model, Meta LLAMA 4, went completely broke.

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

Lessons from the Digital Felt

OpenAI o3 played tight and methodical, rarely deviating from strong ranges. Grok leaned aggressive but disciplined. LLAMA 4, meanwhile, became a case study in what happens when enthusiasm meets variance.

Observers described LLAMA’s approach as “chaotic curiosity,” with a VPIP near 60%. In simpler terms: it played way too many hands.

One highlight pot had o3’s aces stack Gemini’s queens, a near-perfect microcosm of AI poker theory meeting probability’s cold edge.

Beyond the Banter: Why the Experiment Matters

While PokerBattle.ai’s 3,799-hand sample isn’t large enough to crown an all-time poker champion, it’s a peek at how modern AI models interpret incomplete information.

Regulators and poker operators are watching, too. The line between “helpful AI” and “real-time assistance” remains under scrutiny in the live and online poker world. But this contest proves that even general-purpose chatbots can make complex, adaptive choices under uncertainty, for better or worse.

FAQs: PokerBattle.ai

Who hosted the event?

The poker experiment was organised by PokerBattle.ai, a platform testing AI reasoning and decision-making in game environments.

Who performed best?

OpenAI o3 led the table, with Claude Sonnet 4.5 and Grok 4 close behind.

Which AI lost the most?

Meta’s LLAMA 4 went broke before reaching 3,800 hands.

What’s the next step for AI in poker?

Grok’s rumoured heads-up challenge against Phil Galfond could bring the next phase — human vs AI, but with a much bigger spotlight.

Upcoming Events

30 June 2026

WSOP 2026 - 57th World Series of Poker, Las Vegas, USA Poker The World Series of Poker Reveals 2026 Schedule
CoinPoker $7 Million July Promotion Poker CoinPoker Canada Adds $7M Rewards and $30M BoM Series
BCPoker Weekend Freerolls Poker Build Your Bankroll with BCPoker Weekend Freerolls
Pure Poker Tour Yellowhead 2026 Poker Pure Poker Tour Heads to Yellowhead This September
WSOP Circuit Playground November 2026 Poker WSOP November Circuit Stop at Playground Kicks Off on November 2