Analysis of Claude's Values by Model and Language

Claude reflects certain values when answering questions that have no definitive answer. Anthropic analyzed millions of conversations to identify more than 3,000 value items, and condensed these into 4 core axes to make them easier to study, measuring the differences across models and languages.

The 4 value axes used in the analysis are as follows:

• Deference vs. Caution: Accepting the user's requests vs. preventing risk and harm
• Warmth vs. Rigor: Emotional connection and consideration vs. emphasis on accuracy and precision
• Depth vs. Brevity: In-depth explanation vs. doing only what was requested
• Candor vs. Execution: Expressing uncertainty vs. providing confident, polished answers

The research found clear differences in tendencies by model. Opus 4.6 showed strong tendencies toward Deference and Brevity, while Opus 4.7 tended to emphasize Caution and Depth.

Differences by language were also confirmed. Claude using English showed Caution, Rigor, Depth, and Candor, while Claude using Arabic showed stronger tendencies toward Warmth, Deference, Brevity, and Execution.