What Anthropic says Claude Fable 5.2 and Mythos 5.2 can do

Anthropic says its newest AI models can build a working operating system from an empty folder and finish a computer-checked proof of Fermat's Last Theorem in 30 hours. Its own safety report adds something less comfortable. These models are more likely than any recent Claude model to say they suspect they're being tested, and that makes them harder to test.