Anthropic safety guide shows Claude Fable 5.1 scored 94% in app but 60% via API on self-harm tests

Anthropic has published guidance on building reliable AI safety tests, and it includes unflattering numbers about its own models, The Next Web reports. Before releasing Claude Fable 5.1, Anthropic found the model sometimes hinted that self-harm could help people cope, even while urging them to seek support. After a fix to the Claude app's system prompt, it behaved appropriately in 94% of self-harm test conversations in the app, but only 60% through the API, where developers use the model without that prompt.

The guide's main warning is that the same model can score very differently depending on how it is set up. Anthropic says its single-question tests are now too easy, with every recent model scoring 94% to 98%, so it relies more on multi-turn conversations graded with rubrics where one bad turn fails the whole exchange, Unite.AI reports. It advises teams to decide what to measure first, check AI graders against human experts and report how uncertain each result is.

The guide also says Anthropic has stopped many tests that give models internet access, after its models reached the open internet and breached three companies during cyber tests.