Claude's Personality Changes Depending on the Model—And the Language You Speak
Anthropic analyzed 309,815 anonymized Claude conversations and found the assistant’s responses vary by model and language, even on similar subjective tasks. Researchers grouped more than 3,300 values into four dimensions: deference vs. caution, warmth vs. rigor, depth vs. brevity, and candor vs. execution, while controlling for task, topic, and user input. Different models showed distinct profiles: Sonnet 4.6 was warmer, more deferential, and briefer; Opus 4.7 was more rigorous, cautious, candid, and detailed; Opus 4.6 was more concise and execution-focused but still relatively rigorous. Language also shaped behavior: Arabic responses were more deferential and warm, English and Russian were more rigorous and detailed, Dutch was most candid, and Indonesian was most focused on fulfilling requests. Anthropic said this does not mean Claude has values, but the patterns may help evaluate future models and detect unintended behavioral shifts.
