“On a legal agent benchmark, Anthropic's Mythos-5 and Fable-5 models scored 13.3%, significantly outperforming GPT-5.5's score of 2.1%.”