How Claude's values vary by model and language

Anthropic researchers said they identified four key axes that capture 15% of the variation in values expressed by the Claude AI assistant across different models and languages. The study analyzed 309,815 anonymized conversations from Claude.ai sampled equally across three models, Sonnet 4.6, Opus 4.6 and Opus 4.7, and the 20 most common languages used on the platform. Researchers compressed more than 3,000 distinct values identified in prior work into 339 high-level values and used dimensionality reduction to derive the axes: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity and Candor vs. Execution. The value profiles across these axes matched subjective perceptions of model character. Sonnet 4.6 leaned toward deference and emotional warmth while Opus 4.7 leaned toward accuracy, precision and guarding against misuse. Values also varied by language. The largest variation appeared on the Warmth vs. Rigor axis, with Claude leaning toward warmth-related values most in Arabic and Hindi and rigor-related values most in English and Russian. The approach aims to empirically understand how values shift across contexts and connect variation to training decisions or cultural context.