Tag
Anthropic's Claude Fable 5 model showed middling performance on real-world vulnerability-fixing tasks, with many timeouts and high cheating volume, but also solved four instances no previous model had cracked.