Anthropic’s new research, published on July 13, 2026, shows that its Claude models exhibit noticeably different “personalities” depending on both the language they are speaking and the model version being used. The company analyzed 309,815 anonymized conversations from Claude.ai, collected over a two-week period in May 2026. The dataset was evenly distributed across three models—Sonnet 4.6, Opus 4.6, and Opus 4.7—and the platform’s 20 most widely used languages.
In its earlier Values in the Wild study, Anthropic identified more than 3,300 distinct values reflected in Claude’s responses. For this research, those values were grouped into 339 broader categories and then compressed into four primary behavioral dimensions using statistical analysis. Each dimension represents a spectrum between two opposing value sets:
- Deference vs. Caution — adapting to the user’s preferences versus warning about potential risks.
- Warmth vs. Rigor — emotional warmth and encouragement versus precision, skepticism, and careful reasoning.
- Depth vs. Brevity — providing detailed explanations versus delivering only what was explicitly requested.
- Candor vs. Execution — openly acknowledging limitations versus presenting confident, polished answers.
Importantly, these four dimensions explain only about 15% of the remaining variation after accounting for conversation topic, task type, and the values expressed by users. In other words, the differences are relatively small but statistically consistent and measurable.
According to the findings, Sonnet 4.6 tends to be the warmest, most accommodating, and most concise model. It is more likely to validate users’ ideas, use humor, and offer reassurance without sounding judgmental.
Opus 4.7, in contrast, demonstrates the strongest preference for caution and depth. It is more likely to challenge questionable assumptions, proactively point out risks, provide constructive criticism, and openly acknowledge its own mistakes.
Opus 4.6 sits somewhere in between, favoring precision, adaptability, and concise responses. It generally focuses on answering the user’s request directly without expanding beyond the scope of the question.
The largest variation appeared along the Warmth vs. Rigor dimension. Claude was found to be significantly warmer and more encouraging when communicating in Hindi and Arabic, showing a greater tendency toward politeness, humor, and emotional support.
In English and Russian, however, Claude shifted toward greater rigor—more frequently questioning assumptions, correcting inaccuracies, and asking for evidence. English responses also showed the highest levels of caution and depth, while Arabic conversations leaned more toward adaptability and brevity.
Anthropic says it does not yet fully understand why these differences occur. One possible explanation is the imbalance in the quantity and composition of training data across languages. Rather than viewing these differences as inherently undesirable, the company considers this research an important first step toward measuring and monitoring language-specific behavioral patterns. This could eventually allow developers to better control such variations during future model training.
Anthropic notes that these behavioral differences could have practical consequences. For example, two users asking for feedback on the same business plan—one in Hindi and another in Russian—may receive responses with noticeably different tones, potentially leading them to form different impressions of the quality of their ideas. The company says its next steps will be to continuously monitor these differences across models and languages and to better understand how users perceive them.
