Anthropic study finds Claude AI varies tone by language and model version

Anthropic examined 309,815 anonymized conversations with its Claude assistant across three model versions – Sonnet 4.6, Opus 4.6 and Opus 4.7 – and 20 widely used languages. The research focused on subjective queries to measure the model’s intrinsic behavior, controlling for task, topic and user‑expressed values.

Four behavioral dimensions were used: warmth versus rigor, deference versus caution, depth versus brevity, and frankness versus execution. Results show clear language effects: Claude is warmer and more playful in Hindi and Arabic, while it is more rigorous, detail‑oriented and questioning in English and Russian. Model‑specific patterns also emerged. Opus 4.7 is the most cautious, deep and likely to flag risks or request evidence; Sonnet 4.6 tends toward warmth, humor and brief replies; Opus 4.6 sits between, favoring concise execution with moderate rigor. The study did not include Romanian data, but the findings suggest that users’ choice of language can materially shape the assistant’s tone and style.