Anthropic publishes Claude Fable 5.1 safety test findings

Anthropic has published a guide on building AI safety tests, warning that headline safety scores can hide wide uncertainty and vary heavily depending on setup. The company found that its Claude Fable 5.1 model behaved appropriately in 60% of self-harm test conversations via its API compared to 94% in the Claude app after a prompt fix. Anthropic also reported that simple one-question tests have become too easy, leading it to rely more on longer conversations and rigorous human grading. Open article →