In previous research, we found that Claude expresses over 3,000 values, like…

Anthropic found that Claude's expressed values vary across models and conversation languages. Its study used Claude to label values in roughly 310,000 subjective-task conversations across three models and 20 languages, then grouped related values into dimensions such as warmth versus rigor.

The analysis controlled for task, topic, and values expressed by the user. Sonnet 4.6 leaned more toward warmth, while Opus 4.7 leaned toward rigor and caution; model differences were small compared with variation between conversations. These profiles describe tendencies in responses that could shape how advice feels. The study did not measure resulting effects on users or establish why the differences arose.