Anthropic Publishes How Claude Handles Crisis Conversations and Reduces Sycophancy

The results show significant generational improvement. On single-turn crisis responses, Claude 4.5 models respond appropriately 98.6-99.3% of the time. On the harder test - course-correcting mid-conversation - Opus 4.5 scores 91%, up from 36% with Opus 4.1. For sycophancy, the 4.5 family scored 70-85% lower than Opus 4.1 and outperforms all frontier models on the open-source Petri benchmark.

Claude.ai requires users to be 18+, with classifiers flagging self-identified minors. Anthropic is developing a new classifier to detect subtler conversational signs of underage users.