Anthropic says Claude’s behavior shifts by model and language

Anthropic says Claude does not respond with the same style across every version of its chatbot, or across every language. For anyone using AI tools at work, school, or in a portfolio-adjacent business, that matters because small differences in tone, caution, and detail can change how useful an answer feels.

In a research report published Monday, Anthropic said its researchers reviewed 309,815 anonymized Claude conversations involving subjective requests, including advice and feedback. The company said it was studying the “values” Claude expressed, meaning the behavioral priorities that showed up in answers, such as whether the assistant was more cautious, direct, warm, brief, or detailed.

Anthropic said it identified more than 3,300 values in the conversations, then grouped them into four broad dimensions: deference versus caution, warmth versus rigor, depth versus brevity, and candor versus execution. In plain English, the company tried to measure whether Claude tends to agree and encourage, push back and warn, explain at length, or focus on getting the task done.

The researchers said they adjusted for the task, topic, and values expressed by the user so they could compare Claude’s own behavior rather than the kinds of questions people asked.

Different Claude models showed different habits

Anthropic said each Claude model had its own behavioral profile. Sonnet 4.6 leaned more toward warmth, deference, and shorter answers, according to the company. That meant it more often affirmed users and used humor or encouragement.

Opus 4.7 showed more rigor, caution, candor, and depth, Anthropic said. The company said that version more often questioned assumptions, laid out reasoning, flagged risks, and admitted limits in its responses.

Opus 4.6 landed closer to a concise and task-focused style, while still showing more rigor than Sonnet, according to Anthropic.

The company said those findings matched how some users and people inside Anthropic already described the models. Anthropic specifically said Claude.ai users had commented that Opus 4.7 hedged more often than other Claude models, meaning it was more likely to qualify its answers instead of stating them flatly.

Language also changed Claude’s responses

Anthropic said Claude’s behavior shifted depending on the language used in the conversation. Arabic responses tended to show more deference, while English responses placed more weight on caution, according to the report.

The company said Claude was warmest in Hindi and Arabic, with more polite, playful, and encouraging wording. English and Russian responses were more rigorous, according to Anthropic, with more frequent challenges to assumptions, corrections, and requests for evidence.

The findings add a practical wrinkle to the way people compare AI models. Benchmarks often focus on whether a model can solve a coding problem, answer a test question, or complete a task. Anthropic’s report argues that style and judgment also vary in measurable ways, even within products from the same company.

For users, that means choosing an AI model is not only about speed or raw capability. Anthropic’s research suggests the same prompt can feel different depending on the version of Claude selected and the language used to ask it.

This story draws on original reporting from Decrypt.