Poker Arena: Multi-Axis Profiling of Strategic Reasoning and Memory in LLMs

In three linesPoker Arena benchmarks seven frontier LLMs on no-limit Texas Hold'em using a three-layer memory architecture and nine cognitive axes (bet-sizing calibration, positional awareness, etc.). Claude Opus 4.6 wins +$15,730 chips but ranks 5th on mean axis score, showing that scalar leaderboards systematically misrank capability structure.

Summary generated by Claude — human-verified