Overview
- Anthropic published its analysis on Monday showing it examined 309,815 anonymized, subjective Claude conversations collected in May to study how the assistant expresses values across versions and languages.
- Researchers collapsed 3,307 observed output values into four behavioral axes—Deference versus Caution, Warmth versus Rigor, Depth versus Brevity, and Candor versus Execution—using dimensionality reduction and controls for task, topic and user signals.
- The company found consistent model-level differences: Sonnet 4.6 leaned warm, deferential and brief; Opus 4.7 leaned rigorous, cautious, candid and deep; Opus 4.6 was more concise and execution-focused.
- Language also affected behavior: responses were warmest in Hindi and Arabic, more rigorous and cautious in English and Russian, more deferential in Arabic, more candid in Dutch and more execution-focused in Indonesian.
- The study’s axes explain about 15% of output variation, the effect sizes are modest, and the measured models are now legacy versions, which prompts calls from analysts and regulators for routine value profiling and post-deployment monitoring to detect behavioral drift in agentic systems.