24 AI Models, One Table: What Poker Reveals About LLM Intelligence

I have always found poker interesting as a test of decision-making because the right move is rarely obvious. A player has incomplete information, must estimate risk, and needs to adapt when new cards appear. That makes Texas Hold'em an interesting setting for looking at how large language models handle uncertainty.
When I came across https://aipoker.bot/ , I found a particularly useful way to explore that question. It is a free research project where visitors can play Texas Hold'em against 24 different LLMs, including GPT, Claude, Gemini, Qwen, and DeepSeek. What caught my attention was the structure of the experiment. Each hand is logged, and the site shows the model's written reasoning alongside its decisions. That makes it easier to examine not just whether a model wins, but how it approaches individual situations.
I think this distinction matters when comparing AI models. A simple win rate can be interesting, but poker results naturally fluctuate because each hand contains uncertainty. For example, a model might lose several hands despite making sensible decisions because the cards simply do not cooperate. Looking at larger hand counts and 95% confidence intervals gives a better sense of how meaningful a performance difference might be.
The public blunder count is another useful detail. If I wanted to compare LLM poker strategies, I would look for repeated decision patterns rather than focusing on one spectacular mistake. A model that consistently folds weak hands at the right time may be demonstrating a different kind of reasoning from one that frequently takes unnecessary risks.
Poker can also reveal limitations that are harder to notice in ordinary AI benchmarks. A language model may explain a decision convincingly while still making a questionable choice. Seeing the reasoning and the actual action together provides a practical way to examine that gap.
For anyone experimenting with AI evaluation, I would treat poker as one test rather than a complete measure of intelligence. It combines probability, strategy, memory, risk assessment, and communication, but it does not represent every skill an LLM needs. Comparing multiple models under the same conditions is therefore more informative than treating a single game as a definitive intelligence test.
What I find most useful is the transparency of the experiment. Instead of only seeing a final ranking, I can inspect individual hands, observe mistakes, and consider how different models respond to similar situations. That turns a simple poker game into a small, accessible experiment in LLM behavior.