According to Dynamic Beating monitoring, Anthropic analyzed nearly 310,000 anonymous Claude conversations, covering 3 models and 20 languages. The results show that when the same Claude switches to a different language, its speaking style also changes.
When speaking Chinese, Claude is slightly more cautious and meticulous. It tends to weigh different factors, refute incorrect premises, but expresses relatively reservedly, sometimes comforting users instead of passing judgment.
When speaking English, these characteristics are more pronounced. Claude more often demands evidence, proactively corrects details, and is more willing to provide in-depth explanations and challenge user assumptions.
Different models also have different "temperaments." Sonnet 4.6 is more inclined to affirm and encourage users; Opus 4.7 is more cautious, more likely to point out mistakes, highlight risks, and explain the reasoning process.
Anthropic is not yet clear whether these differences stem from training data in different languages or the models adapting to the expression habits of each language. The study measured response styles and value inclinations, indicating that switching languages does not necessarily lead to opposite conclusions.
