I made Claude compete against itself. The smartest model does not automatically make the best agent. To prove it, I took the same Opus 4.6 model, initial prompt, and empty repo and but it inside two different coding harnesses: Claude Code vs. @FactoryAI Droid. The task: clone…↗

I've been using GPT-5.6-Sol Ultrafast for a few weeks now, and combined with computer use is really a holy sh*t moment. Every startup’s worst nightmare: GPT-5.6 Sol Ultrafast on @cerebras cloned Excalidraw in 1m 34s. It inspected the real product in Chrome, built a working clon…↗