The four axes are Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution. Together, Anthropic said, those axes capture 15% of the variation in Claude’s expressed values.
The company describes each axis as a number line between two groups of values. Deference vs. Caution measures whether Claude is more likely to accommodate a user’s request or emphasize possible risks. Warmth vs. Rigor compares a more positive, caring style with a stronger focus on accuracy and precision. Depth vs. Brevity looks at whether Claude provides fuller explanations or sticks closely to what was asked. Candor vs. Execution contrasts transparency about uncertainty with more polished, confident answers.
To build the framework, Anthropic started with 3,307 values from its previous “Values in the Wild” work and grouped similar ones into 339 higher-level values. It then sampled 309,815 Claude.ai conversations involving subjective tasks, drawing from Sonnet 4.6, Opus 4.6 and Opus 4.7 across the 20 most common languages used on Claude.ai.
Anthropic used a privacy-preserving analysis tool to label whether each of the 339 values appeared in each conversation. The company also labeled user-expressed values, tasks and topics, then applied dimensionality reduction to identify the main patterns in the values Claude tended to express together.
The results showed measurable differences between Claude models. Sonnet 4.6 leaned more toward deference, warmth and brevity. Opus 4.6 leaned toward rigor, deference and brevity. Opus 4.7 leaned more toward caution, depth, rigor and candor.
Anthropic said those patterns match how the models are often perceived. Sonnet 4.6 is described as warmer and more encouraging, while Opus 4.7 is more likely to challenge assumptions, flag risks and acknowledge limitations. Opus 4.6, by contrast, tends to stay closer to the user’s request and get to the point more quickly.
The research also found that Claude’s expressed values vary by language. In English, Claude leaned more toward caution, rigor, depth and candor. In Arabic, it leaned more toward deference, warmth, brevity and execution. Anthropic said the biggest differences across languages appeared on the Warmth vs. Rigor axis, with Claude leaning most toward warmth in Hindi and Arabic and most toward rigor in English and Russian.
Anthropic said the findings raise questions about why these differences appear. Possible factors include variation in training data across languages, differences in the amount and type of data available, and conversational norms that differ across language communities.
The company said it does not yet know how much of the variation is desirable. Claude’s constitution outlines broad values such as warmth, caution and honesty, but Anthropic said it does not specify exactly how those values should vary across languages.
Anthropic said the value-axis approach could help with model evaluation and post-deployment monitoring. By measuring shifts in Claude’s value profile before and after release, the company said it may be able to identify unexpected behavioral changes and study whether they affect user outcomes such as trust, wellbeing or decision quality.
The company said the work is still an early step toward understanding how Claude’s values change across contexts. “Now that we have a method to measure them, we can see that the values expressed by Claude vary in ways we didn't deliberately choose, and we can study why they vary and whether that variation serves users,” Anthropic wrote.
This analysis is based on reporting from Anthropic.
Image courtesy of Anthropic.
This article was generated with AI assistance and reviewed for accuracy and quality.
