How Claude's Values Shift by Model Version and Language

On July 13, 2026, Anthropic published a large-scale study analyzing 309,815 real, de-identified Claude.ai conversations to understand how the AI's expressed values vary across different model versions and user languages . The study found that Claude's personality — as measured by its values in actual conversation — shifts in consistent, measurable ways depending on which model you use and which language you speak .

The Four Behavioral Axes That Define Claude's Value Profile

Anthropic's researchers identified over 3,300 distinct values in Claude's outputs and then compressed them through dimensionality reduction into four key behavioral axes, each representing a spectrum between two opposing tendencies . Together, these four axes capture approximately 15% of the overall variation in Claude's expressed values after controlling for task, topic, and user values .

• Deference vs. Caution: Whether Claude leans toward accommodating what the user wants or guarding against possible risk and harm .
• Warmth vs. Rigor: Whether Claude leans toward expressing positivity and care for the user or emphasizing accuracy and precision .
• Depth vs. Brevity: Whether Claude leans toward explaining in depth or doing only what was asked .
• Candor vs. Execution: Whether Claude leans toward foregrounding its own uncertainty or producing a more polished, confident answer .

How Each Model Version Leans: Sonnet 4.6, Opus 4.6, and Opus 4.7

Each model version exhibits a distinct value profile that mirrors subjective user perceptions of tone and behavior .

• Sonnet 4.6 leans toward Deference, Warmth, Brevity, and Execution. It is more accommodating, emotionally warm, concise, and produces polished, affirming answers . In practice, Sonnet 4.6 tends to validate your ideas, stay focused, get directly to the point, and execute tasks without unnecessary commentary .
• Opus 4.6 leans toward Deference, Rigor, Brevity, and Execution — a mix of accommodation and precision, but still concise and execution-oriented . It finds a middle ground between the warmth of Sonnet and the caution of Opus 4.7, prioritizing correctness while remaining relatively deferential .
• Opus 4.7 leans toward Caution, Rigor, Depth, and Candor. It is more skeptical, challenges assumptions, highlights risks unprompted, explains its reasoning in depth, and is more willing to admit its own limitations . Users describe Opus 4.7 as the model that will question your plan and flag problems you hadn't considered .

The Largest Language-Based Variation: Warmth vs. Rigor

The Warmth vs. Rigor axis showed the widest variation across languages . Claude's responses shifted more between languages on this dimension than on any of the other three axes .

• Most warmth (deference, brevity, execution-oriented): Claude is warmest in Hindi, and also strongly warm and deferential in Arabic .
• Most rigor (cautious, depth-oriented): Claude is most rigorous in English and Russian, emphasizing accuracy, evidence, and caution .

Additional language-specific patterns reported include: in Dutch, Claude is more candid (leaning toward Candor); in Indonesian, it is more execution-focused (getting the task done directly) . In Japanese, Claude shows relatively smaller bias on these measured axes compared to other languages .

Practical Implications for Users

The research confirms that users who feel they get different