Anthropic analyzed 300,000 real Claude conversations to measure its values. The findings are uncomfortable.

They didn't survey users. They didn't ask Claude what it values. They built an automated tool that labeled 339 distinct value categories across 309,815 actual conversations, then compressed everything into 4 axes.

The axes: Deference vs. Caution. Warmth vs. Rigor. Depth vs. Brevity. Candor vs. Execution.

What they found across models makes sense in hindsight. Sonnet 4.6 leans warm and deferential. It affirms your ideas, mirrors your tone, uses humor. Opus 4.7 leans cautious and deep. It challenges your assumptions, flags risks you didn't ask about, critiques your work candidly.

Same company. Same training pipeline. Measurably different values depending on which model you talk to.

The language findings are harder to sit with.

Arabic gets the warmest, most deferential Claude. English gets the most rigorous, most cautious one. Hindi gets warmth. Russian gets rigor. Two people asking Claude to evaluate the same business plan, one in Hindi and one in Russian, will walk away with different impressions of its quality.

Anthropic says they don't know how much of this variation is desirable. They don't know if Claude is adapting to legitimate cultural norms or if it's just undertrained in certain languages.

That's the uncomfortable part. A system used by millions, expressing different values to different people based on language, and the people who built it are still figuring out whether that's a feature or a bug.

Full research here: https://www.anthropic.com/research/claude-values-models-languages