OpenRouter's Agent Battle Royale Reveals Failure Modes Invisible to Static Benchmarks

OpenRouter ran a 30-game tournament across eleven language models on 17 June 2026, tracking $482 in inference costs to measure how agents perform under multi-round competitive pressure. The format—elimination across iterative rounds—surfaces different failure modes than MMLU or perplexity scores. Agents must reason, adapt to opponents, and survive successive rounds, not merely produce isolated correct answers. Cost transparency ($16 per game) makes the methodology reproducible for mid-sized teams evaluating production agentic workloads.