#1Claude Mythos Preview Shatters SWE-Bench at 93.9% — Then Gets Locked Away
Anthropic's Claude Mythos Preview set a new world record on SWE-bench Verified at 93.9%, resolving real GitHub issues nearly 19 out of 20 times, while also leading GPQA Diamond at 94.6% and SWE-bench Pro at 77.8%. There's a catch: Anthropic has explicitly said Mythos Preview will not be released for general availability, citing its advanced cybersecurity capabilities. The most powerful coding AI on the planet is also the most restricted.






