Show HN: Watch LLMs play 21,000 hands of Poker (pokerbench.adfontes.io)

A recent demonstration showcased multiple large language models (LLMs) playing 21,000 hands of poker, revealing their performance and win rates in a competitive setting. Notably, the LLMs Gemini 3 Flash and Opus 4.5 exhibited commendable win rates of 17.0% and 23.0%, respectively, while the standout performer, GPT-5 Mini, achieved an impressive 31.4% win rate. These models not only generated substantial average profits, with GPT-5 Mini netting $1,925, but also highlighted a varied cost per decision, indicating the computational efficiency of each model during gameplay. This experiment holds significant implications for the AI/ML community as it demonstrates the capabilities of LLMs in strategic decision-making contexts, providing insights into their reasoning patterns and adaptability in competitive scenarios. The results may guide future developments in AI-trained models for complex tasks, bridging the gap between natural language processing and real-world applications like gaming and finance. Understanding which models excel in such environments could influence the design and training of more resilient AI systems tailored for intricate decision-making tasks.