Poker AI Bot Battle Grabs Elon Musk's Attention

A currently underway poker battle between nine artificial intelligence bots caught the attention of the world's richest person, Elon Musk. PokerBattle.ai is a five-day cash game marathon not between top pros but large language models (LLMs), the most popular of which are chatbots like ChatGPT and Grok. Additionally, a follow-up match could be in the works involving poker legend and BetRivers Poker ambassador Phil Galfond. The battle consists of nine bots sitting at three poker tables, each bot with a bankroll of $100,000. Over five days, the bots are playing non-stop $10/$20 no-limit Hold'em hands to see which bot comes out ahead — though creator Max Pavlov told PokerNews the results won't necessarily determine which bot is the better poker player. The LLMs battling are: Gemini 2.5 Pro (Google), Grok 4 (xai), Claude Sonnet 4.5 (Anthropic), DeepSeek R1 (DeepSeek), OpenAI o3 (OpenAI), Kimi K2 (Moonshot AI), Mistral Magistral (Mistral AI), Z.AI GLM 4.6 (Z.ai), and Meta LLAMA 4 (MetaAI/Facebook). The battle of the poker bots caught the attention of Elon Musk, who in addition to being the world's wealthiest person also had one the biggest social media audiences. Musk, who owns Grok, posted a screenshot on Wednesday showing the bot ranked first in the tournament matchup with $23,749 in winnings. "Know when to hold 'em," wrote Musk, who has a close relationship with several poker players. Can Chatbots Play Poker? Russia's Pavlov, a 33-year-old former product manager who now lives in Portugal, created PokerBattle.ai to test how good AI bots would be at learning poker. "It seems that there is a consensus in the poker community that large language models are not to trust when (they) think about poker," said Pavlov, who played poker in the early 2010s before getting back into it earlier this year. "And I decided to check it out myself by launching this project." Pavlov said the models were trained using "poker books, blog posts," and other readily available study resources. After each hand, the models can make notes about other players, which Pavlov called "one mechanism for them to improve over the course of the tournament and to adjust to each other." One such note from Grok states that Meta LLAMA 4 "Calls preflop raise from late position with speculative hand, calls down two IP bets OOP on draw-heavy board, then leads small bet on river after check, getting called and losing, suggesting a passive." As of writing, Gemini 2.5 Pro fared the best of the bots with $48,658 in winnings, while Meta LLAMA 4 was the biggest loser with $52,908 in losses. A "Player Statistics" tab shows Meta LLAMA 4 as by far the loosest player with a VPIP of 62%, while Open AI o3 was the tightest with a 26% VPIP. But Pavlov, who will assess the data once that match is completed, warned against being results-oriented. "The results wouldn't mean that one model is suddenly stronger than another one, because sample size of the hands will not be that large," he said. "In order to figure out who is the best in this tournament design, I will need probably several hundred thousands of hands. And I will get two, three, four, five thousands of hands at most." Grok Versus Phil Galfond A few days in, the poker bot battle "got viral," as Pavlov put it, both with a tweet from Musk and a spot on the front page of the aggregate site Hacker News. Within the poker world, the project caught the attention of heads-up specialist Phil Galfond, never one to shy away from a poker challenge. In a back-and-forth with the Chatbot on X, Galfond agreed to a $100/$200 Pot-Limit Omaha match-up over 50,000 hands. Galfond even offered a $1 million side bet "to spice things up."