Anthropic BioMysteryBench: Claude Mythos Solves 30% of Expert-Unsolvable Biology Problems
Anthropic·high signal
Anthropic published BioMysteryBench, a 99-question benchmark built from real bioinformatics datasets. Claude Mythos Preview solved 30% of the 23 problems a panel of domain experts could not crack, while Sonnet 4.6 performs on par with experts overall. However, 44% of Mythos's hardest-problem wins are 'brittle' — reproduced fewer than 2 of 5 attempts — suggesting lucky solution paths rather than reproducible strategies.