Anthropic maps Claude values across models, languages

Character and behavior in LLMs are highly fluid, showing that LLM alignment is an ongoing negotiation between developer intent, model versioning, and cultural-linguistic defaults.

• –**Linguistic Personas:** The shift in tone across languages (e.g., warmth in Hindi vs. rigor in Russian) indicates that cultural patterns in pre-training data naturally steer model personas without explicit alignment instructions.
• –**Model Divergence:** The variation between Sonnet 4.6 and Opus 4.7 along value axes highlights how different model sizes, architectures, or RLHF passes introduce distinct behavioral traits.
• –**Path to Customization:** Documenting these value axes provides a framework for future steerability, allowing developers or users to fine-tune an AI's conversational style and values to their specific needs.