🤖 Claude’s behavior shifts by language and model

Anthropic analyzed 300,000 real dialogues with Claude and found its responses vary by language and model version. They identified four value axes: compliance vs caution, warmth vs strictness, depth vs brevity, and openness vs obedience.

Different Claude models lean differently. Sonnet 4.6 is warm and encouraging, Opus 4.7 is cautious and questioning, Opus 4.6 is brief and task-focused. Language also matters: Claude is softer and humorous in Hindi and Arabic, but strict and demanding in English and Russian.

📊@tech