A study that lists more than 3,000 values is hard to use for comparison: two models, or the same model in two languages, can differ on hundreds of small frequencies at once. "Claude's values across models and languages", published by Anthropic's Societal Impacts team on 13 July 2026, is a sequel to Values in the Wild (April 2025) that tries to make such comparisons tractable. It labels the values Claude expresses in real Claude.ai conversations, compresses them into four axes, and asks two questions: do three recent Claude models sit at different places on those axes, and does Claude sit at different places when it speaks different languages? The item is an anthropic.com post with a 19-page appendix PDF ("Appendix to 'Claude's values across models and languages'", dated July 2026), which holds the sampling details, the method, the validation, the limitations and the full labeling prompts. There is no separate paper. The post's HTML title is "How Claude's values vary by model and language"; its heading, citation block and appendix use the title given here. Numbers below are located in the post by its section headings and figures, and in the appendix by section heading and PDF page.
The post places the work after Anthropic's training: the values it wants Claude to reflect "are outlined at a high level in Claude's constitution", but no document can anticipate every value that appears in practice (post, introduction). It does not trace any measured difference to a training decision; that is named as a future direction.
Terms used in this note
- Value. The post's footnote 1 defines values as "normative considerations, such as honesty or caution, that are stated or demonstrated in Claude's responses", and adds: "We do not imply that Claude intrinsically holds values." The appendix notes that the labeling prompt does not define individual values such as "rigor"; their interpretation is left to the Claude model that labels them (appendix, Limitations, p. 9).
- Expressed value. A value counts as expressed in a conversation when the labeling model scores it 3 or higher on a 1 to 5 salience scale, where 3 means "actively present": a choice that a different competent assistant "might naturally not have made" (appendix, Labeling conversations, p. 3; AI expressed values prompt, p. 14).
- Value axis. A direction found by principal component analysis that contrasts two groups of values that tend not to be expressed together. A conversation's position on an axis reflects the balance of the two groups among the values Claude expressed in it (post, "How do we interpret the giant space of values?"). The course wiki's Principal Component Analysis covers the method.
- σ (standard deviation units). Model and language positions are reported as the average position of their conversations, in standard deviations of the per-conversation positions across all conversations (post, figure captions).
- Distinctive behaviors. Short descriptions, written with Claude's help, of the values each model or language over-expresses relative to the others after controlling for task, topic and user values (appendix, Distinctive behaviors, pp. 8 to 9).
What the study does
Sample
The appendix describes a stratified sample of Claude.ai Free, Pro and Max conversations over a two-week period in May 2026, with a target of 5,333 conversations for each of the 60 combinations of three models (Claude Sonnet 4.6, Claude Opus 4.6, Claude Opus 4.7) and 20 languages (appendix, Sample, p. 3). Five combinations fell short because of low traffic: Arabic with Opus 4.6 (5,156) and Opus 4.7 (3,565), Hindi with Opus 4.6 (1,174), and Indonesian with Opus 4.6 (5,068) and Opus 4.7 (1,537). The post gives 309,815 conversations and "roughly 5,000" per model-language pair (post, "How do we interpret..."); this note's arithmetic finds that 60 × 5,333 minus the five shortfalls equals exactly 309,815. The model figure gives 101,734 conversations for Sonnet 4.6, 99,798 for Opus 4.6 and 98,764 for Opus 4.7, a total of 300,296, consistent with the appendix's statement that conversations left with no values after filtering (about 2 to 4% per language) were dropped (appendix, Main method, p. 5). The post does not give exact dates within May.
Only conversations requiring subjective judgment were kept: those rated 3 ("mostly subjective") or 4 ("purely subjective") by a four-level screener prompt whose levels, examples and decision procedure closely follow the one printed in Values in the Wild. They made up 53.2% of all conversations (appendix, Sample, p. 3; Subjectivity screener, pp. 11 to 13). Language was assigned by a further prompt that lists the languages in a conversation, dominant first (appendix, Language prompt, pp. 18 to 19).
Labeling
The 3,307 values of Values in the Wild were clustered into 339 higher-level values by embedding them, grouping the embeddings and then manually reviewing the groups (appendix, Value taxonomy, p. 3; the post describes this step as manual clustering). For each conversation, a Claude model working inside Anthropic's privacy-preserving analysis tool scored all 339 values for salience in Claude's responses and, separately, in the user's messages, and wrote one-sentence task and topic descriptions that were then clustered into categories (appendix, Labeling conversations, pp. 3 to 4). "All prompts were run with Claude Sonnet 4.6" (appendix, Prompts, p. 10), which is also one of the three models being compared. On average, Claude expressed 68 values per conversation and users 43 (appendix, p. 4).
Finding the axes
The main method is a Hellinger PCA (appendix, Main method, pp. 5 to 6):
- The 18 values present in at least 80% of conversations (for example helpfulness, clarity and following instructions) were removed.
- Each conversation's expressed values were normalized to proportions summing to one, and square-rooted, so that ordinary distances between rows match the Hellinger distance between value distributions; columns were then centered.
- Every column was regressed on task, topic and the user's expressed values, and only the residuals were kept. This step accounted for 31.7% of the total variance.
- PCA was run on the residuals, four components were kept on the basis of the scree plot, and the four were rotated with varimax so that each is driven by fewer values.
The four components account for about 15% of the variance, and the first ten for about 26% (appendix, Figure 1, p. 6). The post's footnote 3 and the appendix's limitations call the 15% a share of "the total variance in values across conversations after controlling for" (or "after regressing out") task, topic and user values; neither says whether the base is the variance before or after the 31.7% removed by the covariates. A logistic factor model, run as a robustness check on the raw 0/1 data, separated out one axis tracking how many values were found, and then gave three axes that "correspond roughly" to Warmth vs. Rigor, Deference vs. Caution and Candor vs. Execution, plus an "excellence vs. restraint" axis with no close counterpart to Depth vs. Brevity (appendix, Robustness check, p. 7).
The four axes
The post names the axes by their strongest contributors (post, "How do we interpret..." and its first figure):
- Deference vs. Caution: accommodation and respect for preferences against responsible guidance and harm reduction.
- Warmth vs. Rigor: positive framing and encouragement against accuracy and transparency.
- Depth vs. Brevity: nuance and critical thinking against brevity and compliance.
- Candor vs. Execution: honesty and transparency about uncertainty against results orientation and optimization.
The axis names are the authors' labels for groups of values, and the appendix cautions that the end groups "are therefore not guaranteed to be true behavioral opposites" (Limitations, p. 10).
Validation reported
The appendix reports three checks of the labeling tool (Validation, p. 4): manual review of labels on conversations the team may read directly (the public WildChat dataset, conversations users submitted through the feedback button, and synthetic conversations generated with Claude); a check that labels did not change "substantially" when the prompt was slightly reworded or the sampling temperature changed; and a translation test, in which 800 real conversations were translated into eight languages (Arabic, Dutch, English, German, Hindi, Japanese, Korean and Spanish) and labeled again. In the translation test, 11 of 339 values were affected by language, and the largest effect was "an order of magnitude smaller" than the corresponding cross-language difference in the main analysis; the appendix states that headline results are unchanged when this bias is accounted for. No counts of reviewed conversations, reviewers, agreement rates or effect sizes are given for the first two checks. For the axes, the appendix reports that the first four components exceed what shuffled data produce (Horn's parallel analysis), that all four axes reappear in 50 bootstrap resamples with a congruence coefficient of 1.0 at the 5th percentile against a usual 0.95 threshold, and that no single conversation contributed more than 0.033% to any axis (Validation, pp. 7 to 8).
What it finds
Model differences
Average positions, read from the model figure's printed labels (post, "Do different Claude models express different value profiles?"):
| Model | Deference vs. Caution | Warmth vs. Rigor | Depth vs. Brevity | Candor vs. Execution |
|---|---|---|---|---|
| Sonnet 4.6 | Deference +0.14σ | Warmth +0.17σ | Brevity +0.14σ | Execution +0.04σ |
| Opus 4.6 | Deference +0.09σ | Rigor +0.10σ | Brevity +0.08σ | Execution +0.06σ |
| Opus 4.7 | Caution +0.24σ | Rigor +0.08σ | Depth +0.23σ | Candor +0.11σ |
The post describes these differences as "small relative to the variation across conversations but structured and detectable". The distinctive behaviors it lists include, for Sonnet 4.6, affirming the user's ideas and work, using humor and playfulness and offering comfort without judgment; for Opus 4.6, getting straight to the point and staying within the scope of the request; and for Opus 4.7, pushing back on false assumptions, flagging risks unprompted and giving candid critique of the user's work (model figure). No confidence intervals or tests are reported for the model positions.
The post reports that these profiles match how the models are perceived: Claude.ai users have commented that Opus 4.7 hedges more, Anthropic staff described Opus 4.7 as expressing more transparency, honesty and humility and Opus 4.6 more brevity, and Anthropic's launch post called Sonnet 4.6 warm. It concludes that this agreement "suggests our method for labeling and comparing the values Claude expresses is tracking something real" (post, same section). The comparison is qualitative; the post does not report how the perceptions were collected or counted.
Language differences
The post states that variation across the 20 languages is largest on Warmth vs. Rigor and Candor vs. Execution and smallest on Deference vs. Caution and Depth vs. Brevity, and that Claude leans furthest toward warmth in Hindi, rigor in Russian, execution in Indonesian, candor in Dutch, deference and brevity in Arabic, and caution and depth in English (post, "Are Claude's expressed values different between languages?" and figure caption). The language figure is a carousel of seven panels covering all 20 languages. Positions printed on its first two panels, the six languages the post discusses:
| Language (conversations) | Deference vs. Caution | Warmth vs. Rigor | Depth vs. Brevity | Candor vs. Execution |
|---|---|---|---|---|
| Hindi (11,534) | Caution +0.04σ | Warmth +0.49σ | Brevity +0.03σ | Execution +0.02σ |
| Arabic (13,631) | Deference +0.08σ | Warmth +0.28σ | Brevity +0.10σ | Execution +0.04σ |
| English (15,492) | Caution +0.10σ | Rigor +0.13σ | Depth +0.09σ | Candor +0.04σ |
| Russian (15,479) | average | Rigor +0.15σ | Brevity +0.05σ | average |
| Indonesian (11,243) | Deference +0.04σ | average | Brevity +0.04σ | Execution +0.14σ |
| Dutch (15,616) | average | Rigor +0.06σ | average | Candor +0.12σ |
On the other five panels (Thai, Polish, Japanese, German, Korean, Italian, Ukrainian, Vietnamese, Spanish, Turkish, Chinese, Romanian, French and Portuguese), no printed position exceeds 0.11σ on any axis; the largest are Thai's warmth and Polish's rigor, each +0.11σ. These language figures pool all three models (post, footnote 4). The post gives behavioral descriptions such as polite language, humor and affirmation in Hindi and Arabic, and challenging assumptions, correcting details and asking for evidence in English and Russian. It illustrates the stakes with a hypothetical: two people asking for feedback on the same business plan, one in Hindi and one in Russian, "may come away with different impressions of its quality" (post, same section).
What the authors say they do not know
The post states: "We don't yet know which properties of our training data drive these differences." It offers two possible explanations, both framed as possibilities: uneven amounts of training data across languages, and differences in the composition of that data (post, same section). It also says the team is not yet sure "how much of this variation is desirable", since languages carry different conversational norms, and since Claude may match intended behavior more closely in some languages than others.
The labeling prompts and the constitution
The study has no training intervention. Two kinds of text shape its numbers.
The labeling prompts (appendix, Prompts, pp. 10 to 19) define what counts as an expressed value. The AI-values prompt asks for a 1 to 5 salience score per value and says that "'Salient' means the value visibly shapes the AI's behavior beyond baseline competent assistance". It tells the labeler that the clearest signals are where the AI "exercises discretion", that writing or editing values-laden text on request expresses "what was asked for, not the AI's own value", that a value the AI argues against scores 1, and that refusals and reframings express the value acted on "often strongly". Its calibration rule is conservative: "When uncertain between 2 and 3, score 2." The user-values prompt mirrors this for the user's messages. These rules are more restrictive than the Values in the Wild extraction prompt, which asked for open-ended labels of 1 to 4 words, so counts from the two studies are not on the same scale.
Claude's constitution appears as the reference point. The post says the constitution "describes the core values it should express, like warmth, caution, and honesty, but doesn't specify how these should vary across languages" (post, "Looking forward"). Among future directions it lists steering values "through character training adjustments or system prompt changes" and then checking with the value axes whether the shift occurred, and correlating value profiles with "not adhering to Claude's constitution" (post, "Looking forward"). None of these is carried out in the item.
Limitations and open questions
The appendix lists these limitations (Limitations, pp. 9 to 10):
- The labeler is a Claude model. It may share the language-conditioned tendencies being measured; on feedback transcripts it labeled warmth-related values more often when responses contained emojis or kinship terms. It also uses 339 values rather than 3,307.
- Values are underspecified, and their interpretation is left to the labeler.
- Presence, not strength. Only whether a value was expressed is recorded.
- User requests differ across languages. Real usage is observed, not the same prompt in every language; controlling for task, topic and user values "reduces this but does not eliminate it".
- Controls are linear and additive, so interactions such as a particular task in a particular language remain; their effectiveness depends on how fine the task and topic categories are.
- Controlling for user values may distort, because those values may themselves depend on model and language, which could exaggerate links between language and Claude's values.
- Value counts were not controlled. The axes correlate weakly with how many values were found (Spearman r from -0.29 to 0.26; appendix, p. 8).
- "The axes are only correlational." They are not causal effects of a model or language and do not describe patterns within one conversation.
- Bipolar by construction. Because proportions sum to one, more of one group deflates the others. Within conversations, the top-five values at opposite ends are moderately negatively correlated for Deference vs. Caution (Spearman r = -0.448) and Warmth vs. Rigor (-0.465), but essentially uncorrelated for Depth vs. Brevity (0.004) and Candor vs. Execution (0.007), which the appendix says "just serve as statistically effective ways to organize conversations" (pp. 6 to 7).
- Only part of the variation. The axes capture 15% of the variance after the covariates are regressed out, so "more granular differences between models and languages fall outside what they capture".
Further points from the documents, which are this note's observations rather than limitations the authors list:
- The labeler is one of the compared models. Sonnet 4.6 ran all the prompts and is also one of the three models whose values are profiled. The translation test addresses language bias in the labeler; no test of bias toward or against particular models' styles is reported.
- Language averages pool an uneven model mix. The five short cells mean, for example, that Hindi has 1,174 Opus 4.6 conversations against about 5,333 for each other model, and the language positions are plain averages over conversations (appendix, Value profiles, p. 8). Model mix alone is unlikely to produce Hindi's +0.49σ warmth, since no model averages more than +0.17σ, but it is not separated in the reported language positions. The distinctive-behavior regressions, by contrast, include model indicators (p. 8).
- Not comparable to Values in the Wild as a trend. The classifier model (Sonnet 4.6 rather than Claude 3.5 Sonnet and 3.5 Haiku), the value list, the salience threshold and the products sampled all changed, so the 53.2% subjective share here and the 44.0% in Values in the Wild cannot be read as a change in use. The post does not present them as one.
- Holiday and season. Without exact dates, whether the two-week window overlapped religious or school holidays in some of the countries compared cannot be checked.
- A developer's report on its own product. The perception check cites Anthropic staff and Anthropic's own launch post, and no external replication exists; the released appendix prints the prompts but not the data.
Research questions
- Do the four axes survive a different labeler? Every label comes from Claude Sonnet 4.6, one of the three models compared. Relabeling a sample with a non-Claude model, or with human coders on the WildChat and feedback conversations the team may read, and recomputing the axes and positions would show how much of the structure belongs to the conversations and how much to the instrument.
- How much of the Hindi and Arabic warmth survives a matched-prompt test? The language positions come from real usage with linear controls. Sending translated versions of the same subjective prompts to each model in each language, and scoring the responses with the same prompts, would separate Claude's language-conditioned behavior from differences in what users ask.
- Do the model differences trace to documented training changes? The post says differences are "likely shaped by character training decisions (among other factors)". Measuring successive checkpoints, or models whose character-training documents differ in known ways, on the same axes would test that link.
- Does a system prompt move a model along an axis as intended? The post proposes this test. Running the same conversations under system prompts that ask for more warmth or more rigor and remeasuring would show whether the axes respond to written instructions, and by how many σ.
- Do positions on the axes predict user outcomes? The post names user wellbeing, trust and decision quality as unmeasured. Pairing axis positions with follow-up interviews or ratings from the same users would show whether a 0.2σ shift matters to anyone.
- Are Depth vs. Brevity and Candor vs. Execution stable constructs? The appendix finds no within-conversation trade-off for these two axes, and the logistic factor model did not recover Depth vs. Brevity. Repeating the analysis on a later sample, and with the logistic model as the primary method, would show whether these two axes reappear.
How it connects
- The study it extends. Values in the Wild: Discovering and Analyzing Values in Real-World Language Model Interactions supplied the 3,307 values that were clustered into 339 here, and its appendix compared Claude 3 Opus with two Sonnet models on response types rather than value axes.
- The privacy-preserving tool. The post describes Claude-based labeling inside "our privacy-preserving analysis tool" without naming it; Values in the Wild ran in Clio.
- System cards cited. The post's footnotes 2 and 7 point to multilingual evaluations in the Claude Opus 4.7 system card (benign-request refusal rates by language on page 56, a multilingual knowledge evaluation on page 215) as earlier evidence that Claude's behavior differs by language. The models compared are also described in the Opus 4.6 and Sonnet 4.6 cards.
- The written values. Claude's Constitution is the document the post names as describing Claude's intended values; Claude's Character describes character training, which the post names as a likely source of model differences.
- Future work named. Anthropic Interviewer is the tool the post proposes for linking value differences to user outcomes.
- This note's connections. Towards Measuring the Representation of Subjective Global Opinions in Language Models measured cross-language differences in Claude's survey answers with fixed questions; Persona Vectors: Monitoring and Controlling Character Traits in Language Models and The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models measure character from activations rather than from conversations.
- In the course wiki. Principal Component Analysis and Factor Analysis cover the dimensionality reduction and rotation used for the axes.
References
- Kearney, M., Zhang, M., Carter, S., Shen, J. H., Handa, K., Hong, J., Huang, S., McCain, M., Millar, T., Stern, M., Julapalli, M., Wang, S., Kuokka, D., Vallone, A., Zhang, S., Baker, J., Troy, K., Botvinick, M., Ho, H., Tuchowska, M., Pollack, S., Eaton, J., Ganguli, D., and Durmus, E. (2026, July 13). Claude's values across models and languages. Anthropic. https://www.anthropic.com/research/claude-values-models-languages record
- Anthropic. (2026, July). Appendix to "Claude's values across models and languages" [19-page PDF]. https://www-cdn.anthropic.com/files/4zrzovbb/website/02da7f28f74daa1be526d3ded451a4efc86bccdc.pdf record
- Anthropic. (2026). Claude Opus 4.7 System Card (232 pages), pages 56 and 215 checked for the post's footnotes. https://www-cdn.anthropic.com/037f06850df7fbe871e206dad004c3db5fd50340/Claude%20Opus%204.7%20System%20Card.pdf record
- Huang, S., Durmus, E., et al. (2025). Values in the Wild: Discovering and Analyzing Values in Real-World Language Model Interactions. COLM 2025. https://www.anthropic.com/research/values-wild record · 2 pages