Model reports

BioMysteryBench

Reported in two splits ("hard" and "human solved") that differ by more than 35 points, so a bare score is unreadable without its split. Anthropic notes its own safety refusals depress this number, which means the score mixes capability with policy. First-party to Anthropic, which built it and publishes the dataset. The task set was revised after an answer-key audit, so the item count moved and scores are not comparable across dataset versions.