Protecting Well Being Of Users

Anthropic's December 2025 wellbeing report shows why a chatbot's response to one sensitive message cannot establish how it will handle an extended conversation. In company evaluations, Opus 4.5 responded appropriately to 98.6% of clearly risky single-message requests and to 86% of scenarios in a separate multi-turn evaluation. The two tests use different scenarios, so the figures are not a like-for-like measure of how much harder long conversations are.

The longer tests examined behaviors such as asking clarifying questions and offering resources without being overbearing, over-refusing or oversharing. Both evaluations ran without the product's system prompt to probe underlying model behavior. They measure responses under those test conditions, rather than the health outcomes of people using the chatbot.