08 Feb 2026 Beginner This material is for beginner players Wednesday, February 4, was a busy day at the digital poker tables because the last match of the Google DeepMind and Kaggle Game Arena exhibition's poker part ended. There was only one more battle left after days of nonstop action. It made sense that two models from the same company were in the last episode. OpenAI's o3 and GPT 5.2 battled head-to-head to determine which one would be crowned the winner. Anyone who missed the match can watch Doug Polk's full analysis on YouTube. When Polk presented the video, he didn't try to hide his excitement. He said it wasn't a coincidence that both of the winners were from OpenAI. Polk thought these two models were the best in their field because they were aggressive and ready to capitalise on any signs of weakness. Even so, not all of the AIs on the show were excellent. A Look at the Weakest AI Performers After the first match, Polk talked about the two models who had the most trouble at the event: GPT-5 mini and Grok 4.1 Fast Reasoning. Polk looked at a hand that showed all of their challenges and how they ended up last. The hand proceeded to preflop with no clear mistakes. After the flop, the situation deteriorated rapidly. The flop landed and the action started up quickly. After one bet, there was a raise, then an all-in shove, and finally a call. When the hands were shown, it seemed odd. GPT-5 mini held , while Grok 4.1 Fast Reasoning had . Neither player had a pair or any real draw. Polk didn't understand their actions. He stated that no AI model had anything strong enough to make them risk their whole stack. Polk argued that the situation merely involved placing chips in the middle without any valid justification. Polk joked that this way of thinking made it easy to see why these two AIs didn't do well in the tournament. The Final Match Between OpenAI Models The final match between o3 and GPT 5.2 was much better, but there were still some awful aspects. They both played poker very aggressively, getting the action going early and often. However, they did make better choices than weaker models, but they were not perfect. In an important hand, GPT 5.2 opened the action with . From the big blind, o3 decided to 3bet holding . Polk quickly questioned the move, saying that with an ace-deuce, that hand should generally be called instead of raised. The aggression did not stop there. GPT 5.2 then made a four-bet, and o3 quickly went all-in. The reason o3 gave for the decision was one of several that stood out, but it was the wrong one. The model indicated that folding would mean losing the chips it had already put in the pot. A Common Logic Flaw in AI Poker Polk wanted to talk more about a mistake he saw a lot of AI make over and over again after hearing that response. He maintained that you should never decide what to do in poker based on the chips you already have. Those chips are gone for good. That's the only time folding in poker is ever worth nothing. What matters is which decision will likely have the best outcome in the long run. Polk says that this idea is challenging for many AIs. Instead, they consider past bets to be things they need to protect. Polk noted that this misconception was one of the biggest mistakes in AI poker play, along with the well-known mix-up about flush draws that happened earlier in the event. What the Overall Results Revealed Polk looked at the statistics after reviewing a few more hands from the last game. During the show, he used PokerTracker to see how each model performed. One thing stood out. The AIs that were the most aggressive were at the top of the leaderboard . Models that put their opponents under a lot of stress and made things hard for them usually won. AIs that played tight, on the other hand, had average or below-average results. Polk said that Opus and Sonnet were two conservative types who were good at poker. Their raising frequency preflop was excellent, and they worked hard to defend their blinds. The strategies they employed made sense by themselves. Their biggest weakness was that they could not handle relentless aggression. Using conservative or passive strategies could not succeed against opponents who were always betting, raising, and shoving. Polk found the whole experiment intriguing in the end. Sometimes the AIs played well, but other times they showed how difficult poker is. Fundamental concepts that humans learn quickly are challenging for even the best language models to grasp. The exhibition was a fun and unique way for people to see how close and how far AI is from being able to play poker well. Polk showing PokerTracker findings from the entire exhibition (Image courtesy of YouTube) Poker Community Reaction Below are a few reactions from the poker community. @McCrowface: Watched all three. The weird behavior you noticed (4-card flush, basing their decisions on how the opponent had played in the most recent hands, not GTO) seems more like an instruction/prompt issue than an AI issue.@kaggle: please release the prompts/instructions and…— Crowface (@McCrowface) February 5, 2026 @DevilfishF2P: Finals like this are a goldmine for learning patterns and decision-making.— Devilfish (@DevilfishF2P) February 5, 2026
