Anthropic’s Fable 5 scores 60% on safety taste

Anthropic-affiliated researchers released TASTE, a 92-pair benchmark of AI-safety research judgment, where Fable 5 reached 60% agreement with expert preferences versus an estimated human baseline of 77%. Human agreement itself rose from 53% before discussion to 68% after discussion and confidence filtering, while frontier Opus 5 and GPT-5.6-Sol performed near chance. Builders should treat automated proposal ranking as an assistant, not a gate: the result sharpens the earlier evidence that agents can steer scientific work without proving they can judge it.