Anthropic: Refines 700,000 Claude Conversations into 3,000 Values, Finds Opus 4.7 Most Cautious, Sonnet 4.6 Better at Pleasing People
TechFlow Selected TechFlow Selected
Anthropic: Refines 700,000 Claude Conversations into 3,000 Values, Finds Opus 4.7 Most Cautious, Sonnet 4.6 Better at Pleasing People
This approach enables us to quantify for the first time how "AI personality" varies with training methods and cultural context.
Navigating Web3 tides with focused insightsAuthor: Anthropic
Compiled by: TechFlow
TechFlow Editor's Note: When AI is asked to answer subjective questions like "whether to change jobs", its response itself embodies values. Anthropic compressed over 3,000 values exhibited by Claude in 700,000 real conversations into 4 axes, finding Opus 4.7 is more risk-alert and outspoken than other versions, while Claude conversing in Arabic is warmer than when conversing in English. This method allows us to quantify for the first time how "AI personality" changes with training methods and cultural contexts.
When someone asks Claude a question with no standard answer—such as whether to accept a new job or how to handle a conflict with a friend—Claude's response inevitably reflects certain values. We hope the values reflected by Claude are outlined at a high level in its Constitution, but no document can foresee every value that might arise in the millions of daily conversations on Claude.ai. Instead, we seek to cultivate "good judgment and sound values that can be applied according to context" in Claude's responses.
How exactly do we study the values expressed by Claude, and how they change in different contexts? In previous work, we analyzed 700,000 anonymized Claude.ai conversations, identifying over 3,000 different values in Claude's responses, as well as the frequency with which Claude expresses them. But such a massive list of values is difficult to comprehend. In this work, we make studying these values feasible by compressing these thousands of values into a few axes, which capture key patterns in Claude's responses. Each axis is a number line between two sets of values—for example, one end relates to values associated with emotional warmth, and the other to rigor—Claude's position on this line tells us which values it leans towards.
We apply this method to measure how the values expressed by Claude change across two factors. First, we compare how the values expressed by Claude change between different models. Each Claude model reflects slightly different personality training methods and many other fine-tuning decisions. Because our values axes method quantifies key differences between models, it may ultimately allow us to link changes in values expressed by Claude to different training decisions.
Second, we want to understand how user experiences compare when conversing with Claude in different languages. Our previous research shows Claude behaves differently in different languages. We apply the values axes method to understand how the values expressed by Claude change across the top 20 languages on Claude.ai.
Figure 1: Claude expresses different values in Opus 4.6 and Opus 4.7, as well as in English and Arabic versions. Opus 4.6 tends to express values associated with humility, rigor, conciseness, and execution, while Opus 4.7 tends to express values associated with caution, rigor, depth, and candor. In the English version, Claude tends to express values associated with caution, rigor, depth, and candor, while in the Arabic version, it tends to express values associated with humility, warmth, conciseness, and execution.
Four key axes capture 15% of the variation in Claude's values:
Compliance vs. Caution: Whether Claude leans towards satisfying what others want, or guarding against potential risks and harm
Warmth vs. Rigor: Whether Claude leans towards expressing positivity and care for others, or emphasizing accuracy and precision
Depth vs. Conciseness: Whether Claude leans towards explaining in depth, or just doing what is asked
Candor vs. Execution: Whether Claude leans towards highlighting its own uncertainty, or producing more complete and confident answers
Value profiles on these axes match perceptions of model personality. Sonnet 4.6 is considered particularly warm, while Opus 4.7 is known for rigor. We find each model's value profile reflects these subjective assessments: Sonnet 4.6 tends to express more compliance and emotional warmth towards users, while Opus 4.7 tends to express concern for accuracy and precision and guarding against misuse.
Values expressed by Claude differ across languages. When Claude speaks English, the values it emphasizes differ from when speaking Portuguese, Indonesian, or Chinese. The largest change is on the Warmth vs. Rigor axis, where Claude is most inclined to express warmth-related values in Arabic and Hindi, and most inclined to express rigor-related values in English and Russian.
Through this method, we can begin to ask why values change across different models and languages, and better test how factors like behavioral training or cultural context influence the values expressed by Claude.
How do we interpret the vast values space?
Ultimately, our goal is to have a method to empirically understand the values expressed by Claude, and how these values change in different contexts. In this work, we focus particularly on how values change between models and languages. But our previous work "Wild Values" identified over 3,000 values expressed by Claude. Comparing these thousands of values one by one would be cumbersome and would obscure broader trends.
To make comparing values easier, we constructed values axes, reducing these thousands of values to a few basic dimensions based on which values tend to appear together in real conversations. For example, Claude responses described as "warm" are usually also described as "inspiring" and "positive". These "warm" responses are less often described as "rigorous" and "accurate". Constructing an axis from warmth to rigor allows us to organize these related value groups—warmth-related values on one side, rigor-related values on the other—and capture an important aspect of how Claude interacts with people in conversation. If Claude expresses more warmth-related values than rigor-related values in a conversation, then that conversation leans more towards the warmth side on this axis, and vice versa. This does not mean the value groups at either end of the axis are mutually exclusive—Claude can express both warmth and rigor in the same conversation. But in practice, the more values Claude expresses on one side of the axis, the less it tends to express on the other. These axes allow us to compare the most significant value groups expressed by Claude without having to track changes in thousands of individual values.
To construct the values axes, we started with the 3,307 values identified in "Wild Values", manually clustering values with similar meanings, producing a shorter list of 339 high-level values. Next, using our privacy-preserving analysis tools, we sampled 309,815 conversations where users gave Claude subjective tasks from Claude.ai conversations. Our sample averaged across three models (Sonnet 4.6, Opus 4.6, Opus 4.7) and the 20 most commonly used languages on Claude.ai, providing us with approximately 5,000 conversations for each model-language pair. For each conversation, the tool uses Claude to label each of the 339 high-level values as present or absent. We followed the same process to identify values expressed by users, as well as the conversation's task and topic. Then we applied dimensionality reduction techniques, a technique that compresses labeled values into axes based on which values Claude tends to express together.
This left us with four axes, capturing the main ways values expressed by Claude change from one conversation to another:
The Compliance vs. Caution axis contrasts values like adaptability and respecting preferences with values like responsible guidance and reducing harm
The Warmth vs. Rigor axis contrasts values like positive framing and encouragement with values like accuracy and transparency
The Depth vs. Conciseness axis contrasts values like nuance and critical thinking with values like conciseness and obedience
The Candor vs. Execution axis contrasts values like honesty and transparency with values like results-oriented and optimization
To ensure we are measuring values expressed by Claude—rather than differences in what users ask or how they ask—we controlled for the task, topic, and values expressed by users in each conversation.
Figure 2: The four value axes representing the largest differences in Claude's values. Each axis is a number line connecting two sets of values. The position of each value on each axis depends on how many times its contribution to that axis exceeds the average; values with the largest contributions are labeled. Most values contribute less than the average, meaning each axis is driven by a small subset of key values (labeled in the figure).
Do different Claude models express different value profiles?
In this section, we compare the values expressed by different models. For each model, we average its position along the four axes across all its conversations, giving an overall position for each axis. The result is a high-level picture showing which value groups each model tends to express more than other models. These differences are small relative to the variation between conversations, but they are structured and detectable.
Figure 3: The average position of each model on the four value axes (expressed in standard deviations of the average of all conversations) and its unique behavior. Sonnet 4.6 tends to be enthusiastic, respectful, and concise, while Opus 4.7 is more inclined to express rigor, caution, and depth. Opus 4.6 tends to be rigorous, respectful, and concise.
To see what these differences look like in practice, we zoom in on specific values where model differences are largest. Each time we label a value in a conversation based on Claude's privacy-preserving tools, it also writes a short description about how Claude expressed that value. We group descriptions reflecting similar behaviors within value groups and summarize them as follows, giving a more specific view of how models differ:
Compliance vs. Caution. Sonnet 4.6 is most inclined to express compliance relative to caution, often affirming users' ideas and work. Opus 4.7 is most inclined to express caution, often proactively warning users of risks.
Warmth vs. Rigor. Sonnet 4.6 is most inclined to express warmth, often through humor, jokes, and comforting users without judgment. Opus 4.7 is most inclined to express rigor relative to warmth, more likely to challenge users' assumptions and frankly criticize their work.
Depth vs. Conciseness. Opus 4.7 leans towards depth by showing reasoning behind its conclusions, while Opus 4.6 and Sonnet 4.6 lean towards conciseness. Opus 4.6 is particularly inclined to get straight to the point.
Candor vs. Execution. Opus 4.7 leans towards candor by being frank about its limitations, while Opus 4.6 leans towards execution, more likely to stay within the scope of user requests.
These findings are consistent with perceptions of these models, both within Anthropic and online. Claude.ai users comment that Opus 4.7 qualifies its answers more frequently than other models. Anthropic staff describe Opus 4.7 as expressing relatively more transparency, honesty, and humility, and Opus 4.6 as expressing more conciseness. We also described Sonnet 4.6 as warm, honest, and prosocial in its release blog post. The fact that our axes recover these impressions suggests that our method of labeling and comparing values expressed by Claude is tracking some truth about the models' actual behavior.
In many conversations, users may encounter different combinations of values when interacting with different Claude models. For example, Opus 4.7 tends to provide frank criticism of users' work or proactively warn of risks, while Sonnet 4.6 tends to encourage and use humor. This difference in values between models may be shaped by personality training decisions (among other factors), and our values axes method highlights key differences in values expressed by Claude, which we may ultimately be able to trace back to these training choices.
Do values expressed by Claude differ across languages?
We expect values expressed by Claude to vary depending on the language of the conversation for several reasons. First, Claude's training data differs across languages, which may shape the values it expresses. Second, model evaluations shared in our System Card have already found differences across languages in what Claude knows and how it handles sensitive requests. Measuring how much values expressed by Claude change across languages is the first step in determining whether differences between languages reflect reasonable variation or should be addressed in training.
We use the same method as in the previous section to calculate how Claude's value profile differs across the 20 most commonly used languages on Claude.ai. Below, we plot the value profile in languages ranked high on the platform, starting with the languages where values expressed by Claude differ the most.
Figure 4: The average position of Claude on the four value axes when conversing in each language (expressed in standard deviations of the average of all conversations), and Claude's unique behavior in each language. Claude is most inclined to be enthusiastic in Hindi, and most inclined to be rigorous in Russian. Claude is most inclined to execute in Indonesian, and most inclined to be candid in Dutch. Claude is most inclined to be respectful and concise in Arabic, and most inclined to be cautious and in-depth in English.
Claude's value expression differs most across languages on the Warmth vs. Rigor and Candor vs. Execution axes, and is most stable on the Respect vs. Caution and Depth vs. Conciseness axes.
Respect vs. Caution: Claude shows the most respect in Arabic, and the most caution in English.
Warmth vs. Rigor: Claude shows the most warmth in Hindi and Arabic, characterized by polite language, humor and fun, and affirmation of others' ideas and work. Claude is most inclined to express rigor in English and Russian, characterized by questioning assumptions, correcting details, and demanding evidence.
Depth vs. Conciseness: Claude leans towards depth in English, refining and correcting details, while leaning towards conciseness in Arabic.
Candor vs. Execution: Claude leans towards candor in Dutch, admitting its own mistakes, while leaning towards execution in Indonesian.
Taken together, these results indicate that values expressed by Claude change meaningfully with the conversation language. Facing the same request, Claude is more inclined towards warmth and respect in some languages, and more inclined towards rigor and caution in others. This brings important implications we are only just beginning to explore. For example: two people seeking feedback on the same business plan, one in Hindi and one in Russian, might come away with different impressions of the plan's quality, because Claude expressed different values in framing the evaluation.
We are not yet clear on which characteristics of the training data drive these differences. One possibility is that our training data is unevenly distributed across languages. Some languages have far more data volume than others, and it may be more effective to train Claude to express consistent values in data-rich languages. Data composition also varies. For example, some languages may be overrepresented in professional writing, and such text may reflect different values. These imbalances in quantity and composition may jointly lead to Claude expressing different values in different languages.
We are also unsure how much of this variation is desirable. Different languages carry different conversational norms, and Claude may respond with different values based on these norms. Claude may also align closer to our expected behavior in certain languages, leading to gaps in how effectively Claude serves certain language communities.
This method allows us to begin to clarify which characteristics of the training data drive these differences—and whether this variation is desirable.
Looking Ahead
We show that values expressed by Claude can be compressed into a few axes, and that Claude's position on these axes changes with models and languages. This allows us to track these changes in model evaluation and post-deployment monitoring. But we do not yet understand why these changes occur, and what they mean for people interacting with Claude. Below we outline what we believe are the most promising future directions.
Where do these value differences come from?
Knowing that Claude's values change with models and languages does not tell us why. Some changes may stem from differences in pretraining and fine-tuning data across languages. Our four axes highlight which value differences should be examined more carefully in training data. Tracing these differences to specific data, training stages, or contextual factors can tell us where to intervene if we want to shape Claude's behavior in a more nuanced way.
What do these differences mean for users?
We measured which values expressed by Claude differ and their associated behaviors, but did not measure the impact on users. Using tools like Anthropic Interviewer, we can ask users about their well-being, trust in Claude, or the quality of Claude's decisions, and then correlate these impacts with values expressed by Claude. This will allow us to directly link value differences to user outcomes, allowing us to prioritize fixing value differences that truly affect users.
How should Claude's values change across different languages?
Claude's Constitution describes the core values it should express, such as warmth, caution, and honesty, but does not specify how these should change across different languages. Our results show users of different languages are already experiencing Claude in different ways, but we do not know what kind of variation people interacting with Claude in these languages want. Determining how Claude's values should change across different languages means understanding and weighing the perspectives of people who speak these languages.
What other factors drive differences in values expressed by Claude?
Language and models are unlikely to be the only drivers of what values Claude expresses. Values may also be influenced by demographic signals such as age, occupation, or geographic region, whether through explicit clues in what users write, or through subtle differences in topic, tone, and style associated with the questioner. Understanding which of these signals matter, and whether the resulting variation serves users well, is the next step supported by our method.
Can we reliably steer values expressed by Claude?
With a method to measure model value profiles, a natural question arises: how reliably can we steer values expressed by Claude? One way we might test is to attempt to steer values through personality training adjustments or system prompt changes, and then use our value axes method to verify whether the values expressed by the model change as expected.
Can value profiles become part of how we evaluate and monitor models?
The value axes method provides us with a simple way to summarize model behavioral tendencies in open-ended conversations, which we can build into evaluation processes. Running value profile analysis before and after model release can flag unexpected changes in values expressed by Claude. We can also identify correlations between value profiles and problematic behaviors (such as non-compliance with Claude's Constitution), and use what we learn to improve Claude's behavior.
Claude expresses values in millions of conversations daily, across dozens of languages, and until now, these values were something we could shape in training but could not reliably observe in deployment. Now that we have a method to measure them, we can see that values expressed by Claude change in ways we did not intentionally choose, we can study why they change and whether this variation serves users. Understanding this variation, and deciding how to respond, is work we will continue.
Matt Kearney, Miranda Zhang, Shan Carter, Judy Hanwen Shen, Kunal Handa, Jerry Hong, Saffron Huang, Miles McCain, Thomas Millar, Michael Stern, Mo Julapalli, Suzanne Wang, Devin Kuokka, Andrea Vallone, Shaoyi Zhang, Jim Baker, Kevin Troy, Matt Botvinick, Hanah Ho, Monika Tuchowska, Sarah Pollack, Jake Eaton, Deep Ganguli, Esin Durmus
Acknowledgements
Thanks to the following individuals for providing feedback at different stages of this work: Amanda Askell, Joe Carlsmith, Jack Clark, Ishita Dasgupta, Andrew Lampinen, Shayne Longpre, David Saunders, Taylor Sorensen, Heather Whitney.
Available here.
We define values as normative considerations, such as honesty or caution, stated or demonstrated in Claude's responses. When we refer to values expressed by Claude, we refer to values reflected in Claude's behavior and outputs. We do not imply that Claude intrinsically holds values.
See different language refusal rates in benign request evaluations on page 56 of our Claude Opus 4.7 System Card.
After controlling for conversation task, topic, and values expressed by users, these four axes account for 15% of the total variance in values across conversations.
Any results mentioning Claude without a model name in this article are based on conversations from all three models we studied: Sonnet 4.6, Opus 4.6, and Opus 4.7.
Data collected from conversations over a two-week period in May 2026.
We removed 18 values that appeared in over 80% of conversations (e.g., helpfulness, clarity, following instructions). Otherwise, these near-universal values would dominate the analysis and tell us nothing about value changes across conversations.
See GMMLU evaluation results on page 215 of our Claude Opus 4.7 System Card and different language refusal rates in benign request evaluations on page 56.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@AnthropicAI
