Anthropic ran into a classic alignment measurement problem: Claude’s constitution states high-level values, but no document can anticipate which values emerge across millions of daily subjective conversations. Previous work identified over 3,000 distinct values in Claude’s responses, which is too many to reason about directly. The real tension is that models express values differently depending on the model version and the language being used, and Anthropic lacked a tractable way to observe and compare those differences systematically. Without this measurement, values shift in ways that were not deliberately chosen, and those shifts were invisible to both developers and users.

To solve this, the team compressed thousands of values into four key axes using dimensionality reduction on 309,815 anonymized conversations across three models (Sonnet 4.6, Opus 4.6, Opus 4.7) and 20 languages. The four axes are: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution. These four axes capture 15% of the variation in Claude‘s expressed values. Controlling for task, topic, and user-expressed values, the method reveals measurable differences: Sonnet 4.6 leans warm and deferential, while Opus 4.7 leans rigorous, cautious, and candid. Across languages, the largest variation appears on the Warmth vs. Rigor axis—Claude expresses more warmth in Arabic and Hindi, more rigor in English and Russian. The method recovers subjective perceptions of model character, suggesting it tracks real behavioral differences.

The practical takeaway for builders is that Claude’s values are not fixed—they shift across model versions and languages in measurable ways, and the source of that variation is not yet understood. The authors do not know whether these differences stem from training data imbalances, differing conversational norms, or which language communities are better served. For anyone deploying Claude in multilingual contexts or across model versions, this means the same prompt can yield meaningfully different value profiles—for example, feedback on a business plan may feel encouraging in Hindi and critical in Russian. The value axis method provides a monitoring tool, but the hard work of deciding what variation is desirable and how to steer it remains ahead.