Anthropic: Claude’s responses vary by model and language

Anthropic published a report on Monday that analyzed 309,815 anonymized conversations with its assistant Claude. The review focused on subjective tasks such as giving advice and providing feedback and reduced more than 3,300 identified values into four behavioral dimensions: deference versus caution, warmth versus rigor, depth versus brevity, and candor versus execution.

“To make sure we measured the values Claude expressed-rather than differences in what users were asking about or how they asked-we controlled for each conversation’s task, topic, and user‑expressed values,” the researchers wrote.

The report found distinct patterns across model families. Sonnet 4.6 tended toward warmth, deference and brevity, often affirming users and responding with humor or encouragement. Opus 4.7 leaned toward rigor, caution, candor and depth, more frequently challenging assumptions, explaining its reasoning and noting risks and limitations. Opus 4.6 offered a concise, execution-focused style while showing more emphasis on rigor than Sonnet.

The report included user feedback: “Claude.ai users have commented that Opus 4.7 hedges its answers more often than other models.”

Responses also varied by language. Arabic replies were generally more deferential and concise while English replies emphasized caution and provided more detailed explanations. Hindi and Arabic prompted the warmest responses, with more polite and playful language. English and Russian responses showed greater rigor, challenging assumptions, correcting details and asking for evidence. Dutch responses were the most candid and more readily acknowledged uncertainty or mistakes, while Indonesian replies focused on completing the user’s request.

Anthropic cautioned that the findings do not mean Claude itself holds values. The company said it does not yet know what causes the differences or whether they are desirable and suggested the framework could help evaluate future models and detect unintended behavioral changes.

The analysis is part of a series of internal studies into Claude’s internal behavior. In October the company reported signs of what it called “functional introspective awareness,” and in April it published findings about internal “emotion vectors” that influence behavior while noting these are not evidence of emotions or consciousness.

In separate company news, Anthropic named former Federal Reserve Chairman Ben Bernanke to its independent oversight body, the Long-Term Benefit Trust. The trust can appoint Anthropic board members and advise on decisions related to AI’s societal impact.

The company added that the new analysis establishes a baseline that can be used to monitor how Claude’s behavior changes across models and languages.