Anthropic bio models solved 30 percent of bioinformatics puzzles that humans couldn’t crack in new benchmark

A fresh benchmark exposed how far models can go

Anthropic released results from BioMysteryBench, a new evaluation set for bioinformatics tasks. Its models reportedly solved about 30% of problems that stumped human scientists, suggesting AI is moving beyond math and coding into biological research workflows. The announcement highlights growing confidence in model-driven discovery, while raising questions about real-world usefulness and benchmarks.