A 30-game battle royale across eleven LLMs, $482 of inference, and one finding that should change how you read model benchmarks.