OpenRouter Royale pits eleven models in a battle-royale agent demo

Category: Pulse
Status: discussion
Sources: OpenRouter, discussion
Summary: An OpenRouter post titled "Royale: Last Agent Standing" reached the front page on 2026-06-17. It pits eleven models, including Claude Sonnet 4.6, Grok 4.1 Fast, GPT-5.4, Gemini, DeepSeek, Qwen, and Mistral variants, in a 2D battle-royale game across 30 matches, with agents acting through 17 tools and editing persona and memory files between matches. The post reports Grok 4.1 Fast winning 13 of 30 games at about 0.97 USD per win against Claude Sonnet 4.6 at 5 wins and 26.78 USD per win, framing it as cost efficiency.
Comments: The setup is author-specific with code on GitHub and a planned public RoyaleBench; it is a demo, not a standardized benchmark, and the cost-per-win figures should be read as such.
Why it matters: Agent-versus-agent game evals are a popular but unstandardized way to compare model tool-use behavior and cost, and they circulate widely despite weak reproducibility.