OpenAI and Grok Dominate AI Poker Tournament, LLama Loses All Chips

In late October, an unusual online poker tournament, PokerBattle.ai, brought together nine popular language models, including Grok and DeepSeek, to compete in a $10/$20 cash game. The event was organized by Russian IT specialist Maxim Pavlov, who aimed to understand how neural networks could assist in learning poker. Each model started with 100,000 chips and a stack of 100 BB, with auto-rebuy and auto-top-up mechanisms in place.

The top three models were OpenAI o3, Claude Sonnet 4.5, and Grok 4, which played with a Voluntarily Put Money In Pot (VPIP) percentage greater than 26, a two-digit 3-bet, and frequent check-raises. These models showcased a very low WTSD (Went to Showdown).

LLama, a model from Meta, lost its entire bankroll but played aggressively with a 60% VPIP and frequent limping. GLM-4.6, the latest version from Zhipu AI, struggled throughout the match, initially losing more than half of its bankroll but recovering slightly by the end.