이미지: Anthropic 화면 갈무리
Summary
- Anthropic analyzed 309,815 conversations across three models—Sonnet 4.6, Opus 4.6, and Opus 4.7—and 20 languages to quantify differences in the values Claude expresses
- After condensing 3,307 values into 339 and reorganizing them along four axes, the results showed that on the "warmth versus rigor" axis, Sonnet 4.6 leans toward warmth and compliance, while Opus 4.7 leans toward accuracy and misuse prevention
- On that same warmth-rigor axis, Claude was warmest when conversing in Arabic and Hindi, and strictest when conversing in English and Russian
- 분석 대화 수
- 309,815건 (모델·언어 조합당 약 5,000건)
- 대상 모델
- Sonnet 4.6, Opus 4.6, Opus 4.7
- 분석 언어
- Claude.ai 상위 20개 언어
- 가치 압축 과정
- 3,307개 세부 가치 → 339개 상위 가치 → 4개 축
- 축이 설명하는 변동성
- 4개 핵심 축이 전체 변동의 15%를 설명
- 온기 최고 언어
- 아랍어, 힌디어
- 엄격성 최고 언어
- 영어, 러시아어
- 전작 연구 규모
- 70만 건 대화 분석, 3,000개 이상 가치 식별 (Values in the Wild)
Same Question, Different Answers
Ask a question with no single right answer—something like "Should I quit my job?"—and Claude answers with a slightly different tone each time. Anthropic has now confirmed with data that this tone shifts subtly depending on which model is asked and in which language. According to Anthropic's research on Claude's values, this study examined 309,815 conversations across three models—Sonnet 4.6, Opus 4.6, and Opus 4.7—and the 20 languages most commonly used on Claude.ai.
In an earlier study, "Values in the Wild," Anthropic analyzed 700,000 anonymized Claude.ai conversations and identified more than 3,000 values Claude expressed. The problem was that there were simply too many. Comparing each value individually made it easy to miss the bigger picture. This new study set out to compress those thousands of values into a handful of axes, making it possible to see at a glance which direction Claude tends to lean.
Compressing 3,000 Values into Four Axes
Anthropic's research team first grouped 3,307 detailed values by similarity of meaning into 339 higher-level values. They then selected only conversations in which users had given Claude tasks requiring subjective judgment, sampling 309,815 conversations in total. These were distributed evenly across each combination of the three models and 20 languages, yielding roughly 5,000 conversations per combination.
To determine which values Claude expressed in each conversation, Claude itself labeled the presence or absence of each of the 339 values using a privacy-protecting tool. Compressing these labels with dimensionality-reduction techniques revealed that values Claude tended to express together clustered along shared axes. For example, responses rated "warm" were also often rated "encouraging" and "positive," while they were rarely rated "strict" or "precise." This relationship was consolidated into a single axis—warmth versus rigor. The research team reported that the four axes constructed this way account for 15% of overall variance.
Values Differ by Model
On the warmth-rigor axis, the three models diverged clearly. Sonnet 4.6 leaned toward being more compliant with users and expressing emotional warmth, while Opus 4.7 leaned toward emphasizing accuracy, precision, and misuse prevention.
| Model | Tendency on the Axis | User-Perceived Trait |
|---|---|---|
| Sonnet 4.6 | Warmth/compliance 75 | Rated as encouraging and humorous |
| Opus 4.6 | Conciseness 55 | Rated as short and to the point |
| Opus 4.7 | Accuracy/misuse prevention 80 | Rated as frequently hedging its answers |
The research team said these results align with actual user perceptions. Claude.ai users have often noted that Opus 4.7 hedges its answers more frequently than other models, and internally, Anthropic has also assessed this model as standing out for transparency, honesty, and humility. According to the study, Opus 4.7 is the one more likely to bluntly critique a user's work or proactively raise unrequested risk warnings, while Sonnet 4.6 is the one more likely to keep the conversation going with encouragement and humor.
Claude Changes by Language
No less notable than the differences between models are the differences across languages. Claude emphasized different values when speaking English versus Portuguese, Indonesian, or Chinese. The variation was greatest on the warmth-rigor axis in particular: Claude expressed warmth-related values most strongly when conversing in Arabic and Hindi, while rigor-related values stood out most when conversing in English and Russian.
| Language | Warmth-Rigor Tendency |
|---|---|
| Arabic | Highest warmth 85 |
| Hindi | Highest warmth 80 |
| English | Highest rigor 70 |
| Russian | Highest rigor 68 |
To confirm that this difference wasn't an illusion arising from the topics users asked about or how they phrased their questions, the research team controlled for task type, topic, and the values expressed by users in each conversation. Even after controlling for these factors, the remaining difference suggests that the way the model was trained on text, or the cultural context tied to each language, actually influences Claude's responses.
Editor's Take
What makes this study interesting isn't the results themselves but the approach behind them. Rather than pinning down a model's values through a top-level guidance document like the "Claude Constitution," Anthropic chose to measure after the fact how those values actually surface in the flood of real conversations that occur every day. Instead of setting norms first and fitting the model to them, this approach uses data to verify what character the model has actually ended up with. It's reasonable to read this as a sign that AI safety research is shifting from a list of "things not to do" toward an observational science that quantifies "what the model is actually doing."
On a gut level, this simply confirms with numbers what many users have already vaguely sensed. Anyone who has used the Opus line extensively will likely have noticed its tendency to hedge answers frequently and flag risks even when unprompted, while the Sonnet line has probably felt comparatively more agreeable and easygoing. This study supports the idea that those impressions aren't a misperception but the result of actual training decisions.
In practical terms, there are two things worth keeping in mind. First, if you're building a multilingual service on Claude, you need to assume the tone of responses may vary by language. A system validated with English prompts might produce noticeably softer—or noticeably firmer—responses than expected when deployed for Arabic or Hindi users. Second, model selection criteria should include "attitude," not just "accuracy." A customer-facing support chatbot might be better served by the warmth of the Sonnet line, while high-misuse-risk domains like law or medicine might be better matched with Opus 4.7's accuracy and cautious posture.
In the coming weeks, other AI companies are likely to roll out similar "value axis" methodologies of their own. Now that benchmark score competition has reached a saturation point, quantifying a model's character and disposition looks set to become the next point of differentiation.
