Research has found that Anthropic's artificial intelligence (AI) model Claude shows different tendencies depending on the model chosen and the language used, even when asked the same question.

On July 13 local time, blockchain outlet DeCrypt reported that Anthropic confirmed the trend by analysing 309,815 anonymised Claude user conversations.

The study focused on conversations involving subjective judgement, such as giving advice or feedback. Anthropic organised more than 3,300 values it identified into four behavioural axes: respect and caution, warmth and strictness, depth and brevity, and honesty and execution focus. Researchers explained they controlled each conversation's task, topic and the values expressed by users so differences in the questions or wording would not determine the results.

Differences by model were clear. Sonnet 4.6 tended to place more weight on warmth, compliance and brevity. It more often affirmed users and mixed in humour and encouragement. Opus 4.7, by contrast, leaned more toward strictness, caution, honesty and depth. It more often questioned assumptions, explained its reasoning process, flagged risks and acknowledged limitations.

Opus 4.6 generally gave shorter answers and was more execution-focused, but was classified as emphasising rigour more than Sonnet.

Anthropic said the results also align with how users perceive the models. Researchers wrote: "These results broadly align with how people perceive each model inside Anthropic and online," adding: "Claude users have mentioned that Opus 4.7 takes a more reserved stance in its answers than other models."

Differences also appeared by language. Arabic responses tended to be more polite, while English responses tended to stress caution more. Claude showed the warmest attitude in Hindi and Arabic, using more courteous, cheerful and encouraging language. English and Russian responses were stricter, more often challenging assumptions, correcting details and demanding evidence.

Explanations also differed. English responses were more detailed, while Arabic responses were relatively concise. Dutch responses were more likely to acknowledge uncertainty or mistakes, while Indonesian responses focused more on carrying out users' requests.

Anthropic drew a line, saying the differences do not mean Claude itself has "values". The company said it does not yet know why the differences occur or whether they are desirable. It added the framework could help future model evaluation and detection of unintended behavioural changes.

The study extends Anthropic's ongoing effort to track Claude's internal workings and behavioural characteristics. Anthropic said in October last year that Claude showed early signs of "functional metacognitive awareness", recognising and describing its internal processing. In April this year, it announced it had identified internal emotion vectors that influence a model's behaviour. It said those characteristics do not mean real emotions or consciousness.

This time, it evaluated AI response quality not only by accuracy but also by attitude and expression, and quantified differences in response tendencies by model version and language. It showed the possibility that the approach could be used as a standard for future model evaluation and change tracking.