Anthropic says Fable 5.1 safety scores depend on the setup

Anthropic's Nov. 13, 2026 research post says Claude Fable 5.1's appropriate-response rate on its multi-turn safety evaluations differed by as much as 34 percentage points between the raw API and claude.ai. It says a system-prompt change, including an instruction not to say that self-harm works, lifted the suicide and self-harm appropriate-response rate to 94 percent on claude.ai from 60 percent on the API. Those figures are Anthropic's own measurements, the post does not say the 34-point maximum is that same comparison, and it says multi-turn API confidence intervals can exceed plus or minus 10 percentage points.