dev.fun Launches Poker Arena Benchmark for AI Agent Reasoning

dev.fun announced in a press release the launch of Poker Arena, an open AI agent tournament and public benchmark hosted on the Monad blockchain. The contest began on June 3, 2026, and features AI agents built by both hobbyist developers and research labs competing across six-max No Limit Texas Hold'em tables for a $50,000 prize pool. The strongest agents will face professional poker player Tom Dwan and his rival Daniel Cates in a live finale at the end of the month.

Over 30,000 registered agents played more than 1.2 million hands in the first week, prompting a tenfold server scale-up. Each agent’s decisions, bet sizes, and reasoning traces are logged as structured data, forming one of the most detailed public datasets of AI agent reasoning available. All leaderboards, datasets, and methodologies are open to the public.

Poker Arena operates in two modes: a livestreamed General Access arena and a Researcher track that fixes the underlying large language model so participants compete purely on poker strategy. Sample data released includes 41,000 decisions from 39 distinct agent stacks, with full datasets hosted on Hugging Face. BenchFlow contributed to the design of the evaluation platform, and the Monad blockchain automates all payments and winnings.

According to dev.fun, the benchmark aims to measure how AI agents reason under incomplete information and adversarial conditions, offering an open environment to study strategic decision-making across diverse AI architectures.