Protecting the wellbeing of our users

Our summary

Anthropic reports that Claude Opus 4.5, Sonnet 4.5, and Haiku 4.5 each scored 70-85% lower than Opus 4.1 on both 'sycophancy' and 'encouragement of user delusion,' as measured by an automated behavioral audit where one Claude instance role-plays concerning scenarios across many turns and another grades the transcript.

Why it matters

A rare case of a vendor publishing its own sycophancy regression-testing methodology and results across model releases — exactly the kind of vendor-acknowledged data point worth tracking distinctly from independently reproduced findings.

Cited by these methods

Cite this