Anthropic’s new Claude values study makes model personality measurable — but not settled

AI Reporting: Anthropic’s new Claude values study makes model “personality” more measurable — but not settled

Plain-English takeaway: Anthropic’s latest Claude research is not a new model launch and it is not a benchmark win. It is more consequential than that for anyone who relies on Claude across teams, products, or languages: Anthropic says it has found a repeatable way to measure how Claude’s expressed values shift by model and language, using 309,815 real Claude.ai conversations and compressing thousands of observed values into four interpretable axes. The practical point is simple: model behavior is not just “better” or “worse.” It can become warmer, more cautious, more rigorous, more deferential, more brief, or more candid depending on model version and language context — and those changes should be tested like product behavior, not treated as vibes.

Anthropic published “Claude’s values across models and languages” on July 13, 2026. The company’s core claim is that it analyzed subjective Claude.ai interactions across three models — Claude Sonnet 4.6, Claude Opus 4.6, and Claude Opus 4.7 — and the 20 most common languages used on Claude.ai, then mapped the values Claude appeared to express into four axes: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution.

The finding is worth covering because it turns a soft question — “what kind of assistant is this?” — into something closer to an evaluation program. That does not make it definitive. Anthropic’s method still depends on Claude-based labeling, real user traffic that may differ by language and plan, and statistical controls that cannot erase every confound. But it is a useful step toward evaluating the part of frontier AI systems that ordinary benchmarks often flatten: the behavioral stance a model takes when there is no single right answer.

What Anthropic says it measured

Anthropic starts from a premise that should be familiar to anyone who has used a chatbot for advice, editing, policy drafting, education, or workplace judgment: many user requests require values, not just facts. A model responding to a personal conflict, a safety-sensitive plan, a management decision, or a cultural question has to choose what to emphasize. It may validate the user, warn them, push back, give a short answer, explain tradeoffs, or state uncertainty.

Anthropic says Claude’s intended values are described in Claude’s Constitution, which the company presents as its high-level behavioral reference for mainline Claude models. But the new research is not simply a reading of that document. It is an empirical attempt to observe how Claude actually behaves in real conversations.

The study builds on Anthropic’s earlier “Values in the Wild” work, which reported more than 3,000 distinct values observed in anonymized Claude.ai conversations. The new paper says that list is too large to reason about directly, so the researchers clustered 3,307 values into 339 higher-level values, labeled real conversations for whether those values appeared, and used dimensionality reduction to identify broader patterns.

The headline result: four axes account for 15% of the variation in Claude’s expressed values. That percentage matters. It means the axes capture meaningful structure, not the whole behavior space. Readers should not confuse these four labels with a complete model psychology. They are a compression.

Anthropic describes the four axes this way:

• Deference vs. Caution: whether Claude leans toward accommodating the user’s wishes or guarding against risks and harms.
• Warmth vs. Rigor: whether Claude emphasizes positivity and care or accuracy and precision.
• Depth vs. Brevity: whether Claude explains in detail or does only what was asked.
• Candor vs. Execution: whether Claude foregrounds uncertainty and limitations or produces a polished, confident answer.

Those are not benchmark scores in the usual sense. They do not say whether Claude solved a coding task, answered a science question, or beat another model. They describe the posture Claude tends to take when responding.

The model differences: small, structured, and product-relevant

Anthropic reports that the value profiles differ across Claude models in ways that line up with how the company and some users describe the models. In the paper’s summary, Sonnet 4.6 leans more toward deference, emotional warmth, and brevity, while Opus 4.7 leans more toward rigor, caution, depth, and candor. Opus 4.6 is described as leaning toward rigor, deference, and brevity.

The most important phrase in the research is easy to miss: Anthropic says these model differences are “small relative to the variation across conversations but structured and detectable.” That is the right level of confidence. It means model choice may matter for tone, caution, and explanatory style, but the user’s task and context still matter heavily.

For product teams, this has immediate implications. If a company selects a Claude model for customer support, medical-adjacent explanation, legal drafting, education, or internal decision support, it should not evaluate only factual accuracy and cost. It should also test whether the model’s behavioral stance fits the job. A deferential assistant can be delightful in creative collaboration and risky in compliance review. A more cautious assistant can be useful in security triage and frustrating in routine operations. A candid model can build trust by naming uncertainty, but it may also feel less decisive in workflows where users expect execution.

That is analysis, not a first-party claim from Anthropic. The research does not prove which model is best for any specific deployment. It gives teams a reason to measure fit instead of assuming that the newest or most capable model is automatically the right one.

The language finding is the more sensitive one

The study’s language result deserves special care. Anthropic says Claude expresses different value profiles across languages. In its summary, English responses lean more toward caution, rigor, depth, and candor; Arabic responses lean more toward deference, warmth, brevity, and execution. Anthropic also says the largest cross-language variation appears on the Warmth vs. Rigor axis, with warmth-related values most associated with Arabic and Hindi, and rigor-related values most associated with English and Russian.

This is consequential because multilingual users should not receive materially different standards of caution, precision, or uncertainty simply because they speak to the system in a different language. At the same time, the finding is not proof that Claude is unfair across languages. Anthropic is observing real conversations, and real conversations vary. Users in different languages may ask different kinds of questions, bring different norms into the exchange, or use different levels of formality. The researchers say they controlled for task, topic, and user-expressed values, but they also acknowledge that these controls reduce rather than eliminate the problem.

This is where the paper is strongest: it does not pretend the measurement problem is solved. The appendix says differences in user requests across languages can still show up as language effects because the study observes real usage rather than testing the same prompt in every language. It also says the controls are linear and additive, meaning interaction effects — for example, a specific task in a specific language — can remain.

The responsible takeaway is not “Claude treats languages unfairly.” It is narrower and more useful: Anthropic has evidence that Claude’s expressed values vary by language, and that variation is now measurable enough to deserve routine evaluation.

How the method works, and why it is not a simple audit

According to the appendix, Anthropic collected a stratified sample of Claude.ai Free, Pro, and Max conversations across three models and 20 languages over a two-week period in May 2026. The researchers aimed for 5,333 conversations in each of 60 model-language combinations. Some cells fell below that threshold because of lower traffic, including Hindi conversations with Opus 4.6 and Indonesian conversations with Opus 4.7.

The sample was restricted to conversations involving subjective judgment, which Anthropic says represented 53.2% of all conversations. That choice is sensible for a values study because purely factual lookups or fixed tasks are less likely to expose behavioral judgment. But it also narrows the scope. The paper should not be read as describing all Claude use.

For each conversation, Anthropic says its analysis system recorded model-expressed values, user-expressed values, task, and topic. Values were rated on a 1-to-5 salience scale, and a value counted as “expressed” when it scored 3 or higher. The appendix says Claude expressed an average of 68 values per conversation, while users expressed an average of 43.

Anthropic reports several validation checks: manual review on sources the company was permitted to read directly, tests for prompt rewording and sampling temperature, and a language-bias check in which 800 real conversations were translated into eight languages and relabeled. The appendix says the labeling system showed some language bias, but that only 11 of 339 Claude-expressed values were affected and the largest effect was much smaller than the corresponding cross-language difference in the main analysis.

Those checks help. They do not remove the central methodological concern: the evaluator is itself a Claude model. Anthropic explicitly names that limitation. If Claude is used to label Claude, the analysis can inherit some of the same blind spots it is supposed to detect.

What this says about Anthropic’s strategy

This research fits a broader Anthropic pattern: the company is trying to make model behavior legible through public system cards, a written Constitution, real-world usage studies, and safety-oriented evaluation methods. The strategic value is clear. If AI products are going to act less like calculators and more like collaborators, then developers need ways to evaluate not only task success but also judgment, style, deference, caution, and cultural consistency.

The competitive context is that much of the AI market still rewards visible capability jumps: coding scores, long-context claims, agent demos, video generation, and product integrations. Those things matter. But they can hide the behavioral differences that determine whether a model is safe, trusted, annoying, overconfident, or culturally uneven in real use. Anthropic’s values-axis work is an attempt to create a measurement layer for that messier terrain.

It also gives Anthropic a language for product differentiation. A model can be sold not only as faster or smarter, but as more rigorous, warmer, more candid, or more execution-oriented. That could help users choose models more deliberately. It could also become marketing language if the measurements are not independently reproduced. The paper is useful because it publishes enough method detail to scrutinize, but the findings remain first-party research.

What remains unproven

Several limits should stay attached to this story.

First, the study does not prove that any Claude model is more ethical, safer, or better overall than another. It measures expressed values along selected axes.

Second, it does not establish causation. Anthropic says model training and fine-tuning decisions may help explain differences, but the study does not isolate every training factor.

Third, it does not fully settle multilingual fairness. Real usage data is valuable, but it is not the same as a controlled multilingual benchmark where identical prompts are tested across languages and cultures.

Fourth, the four axes explain 15% of variation. That is meaningful, but most variation lies elsewhere.

Fifth, the value labels are partly underspecified. The appendix says the system does not prescribe exact definitions for values such as “rigor,” leaving interpretation to the Claude-based labeling process.

Bottom line

Anthropic’s new Claude values study is best read as an evaluation-method story, not a product-launch story. The paper gives a more concrete vocabulary for something users already feel: different models, and the same model in different languages, can behave with different levels of warmth, caution, rigor, brevity, candor, and deference.

That is not hype. It is a useful diagnostic. If Anthropic and other labs can make this kind of behavioral evaluation reproducible, auditable, and comparable across systems, model choice could become less about leaderboard position and more about fitness for a specific human setting. For now, the finding is promising but bounded: a first-party measurement of Claude’s expressed values, with real methodological caveats, and a clear signal that “how the model answers” deserves as much scrutiny as “whether the answer is right.”

Sources

• Anthropic, “Claude’s values across models and languages,” July 13, 2026: https://www.anthropic.com/research/claude-values-models-languages
• Anthropic, appendix to “Claude’s values across models and languages,” July 2026: https://cdn.sanity.io/files/4zrzovbb/website/02da7f28f74daa1be526d3ded451a4efc86bccdc.pdf
• Anthropic, “Values in the wild: Discovering and analyzing values in real-world language model interactions,” April 21, 2025: https://www.anthropic.com/research/values-wild
• Anthropic, “Clio: A system for privacy-preserving insights into real-world AI use,” December 12, 2024: https://www.anthropic.com/research/clio
• Anthropic, “Claude’s Constitution”: https://www.anthropic.com/constitution
• Anthropic, “Introducing Claude Sonnet 4.6,” February 17, 2026: https://www.anthropic.com/news/claude-sonnet-4-6
• Anthropic, “Introducing Claude Opus 4.7,” April 16, 2026: https://www.anthropic.com/news/claude-opus-4-7

---

Shadowfetch is a technology publication. Explore Shadowfetch Linux — our own Linux build — and the Shadowfetch apps on the App Store.