Even GPT-5.6 Can't Simulate Physics

🔥 What happened Researchers dropped WorldSolver: 168 physics simulation tasks pulled from 61 classic graphics papers. LLM agents must write solver code that reproduces real physical dynamics across 7 domains. 💡 Why it matters Even frontier models flop hard: GPT-5.6-Sol scores just 48.7%, Claude-Opus-5 hits 46.7%. Getting code to run is hard — getting the physics right is brutal. ⚡ Our take If you think LLMs understand physics, read this benchmark. They fail half the tasks — and that's with a pre-built scaffold handed to them.