Claude value profiles diverge across 300K chats, models and languages

Claude gets a measurable value profile

Anthropic has turned a vague product question into a measurable research problem: does Claude express different values depending on the model or language being used? In a July 13 X post, the company wrote, “We analyzed 300K+ anonymized conversations,” linking to a new study of Claude’s behavior across models and languages. The source tweet is available here.

The research matters because multilingual AI quality is not only about factual accuracy or translation fluency. A model can answer correctly while still sounding more deferential, cautious, warm, brief, candid, or execution-focused depending on the context. Anthropic’s team started from more than 3,000 values identified in prior work, compressed them into higher-level groups, and analyzed 309,815 anonymized Claude.ai conversations. The sample covered Sonnet 4.6, Opus 4.6, Opus 4.7, and the top 20 languages used on Claude.ai.

The resulting framework uses four axes: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution. Anthropic says those axes capture the main ways Claude’s expressed values shift. Sonnet 4.6 leaned more toward warmth, deference, and brevity, while Opus 4.7 leaned toward caution, depth, rigor, and candid critique. Across languages, the largest contrast appeared on the Warmth vs. Rigor axis: Claude leaned warmer in Hindi and Arabic, and more rigorous in Russian and English.

Anthropic usually uses its X account for model launches, safety research, and interpretability work; this post fits the safety and evaluation side of that pattern. The linked research article argues that value profiling could help connect behavior changes to training decisions, system prompts, or data composition. It also raises a harder question: whether users in different language communities want the same conversational values from an assistant.

What to watch next is whether this becomes an operational evaluation rather than a one-off analysis. If value profiles are run before launch and after deployment, labs could catch unintended shifts in model character the same way they now monitor benchmark scores, refusals, and safety incidents.