A robot is sprinting towards you. Do you want it running on Claude or Grok?

A developer tested 11 large language models in a custom 2D battle royale game over 30 matches to evaluate their real-world performance beyond standard benchmarks. Grok 4.1 Fast won the most games (13 wins) at the lowest cost per win ($0.97), while Claude Sonnet 4.6 prioritized cooperation and communication despite fewer wins, demonstrating that competitive gaming performance doesn't necessarily reflect practical utility. The results revealed that traditional AI benchmarks failed to predict which models would succeed in this dynamic, strategic task.