Claude’s expressed values vary in measurable ways across models and languages, but the differences are averages—not rules for every answer. Anthropic’s July 2026 study found tendencies toward different response styles in three Claude models and 20 languages. Its four summary axes capture 15% of value-pattern variation after controlling for task, topic, and values expressed by users. The study measures outputs, not beliefs, and does not establish what causes the differences or how they affect users.
What Anthropic means by Claude’s “values”
Anthropic uses “values” to mean normative considerations—such as honesty or caution—that Claude states or demonstrates in responses. The term describes observed behavior, not an assertion that Claude intrinsically holds beliefs. As the authors put it, “We do not imply that Claude intrinsically holds values.” (Anthropic, July 13, 2026.)
![Claude AI for Beginners Bible: [5 in 1] The Ultimate Guide to Automate Your Work, Save Hours Every Week, and Use AI for Real-World Results](https://m.media-amazon.com/images/I/415+fSJacsL._SL160_.jpg)
As an Amazon Associate I earn from qualifying purchases.
The study began with 3,307 values identified in earlier research, then manually grouped similar items into 339 high-level categories. Researchers used dimensionality reduction to summarize how those values co-occurred in conversations. The earlier set comes from Anthropic’s Values in the Wild work.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallHow the study was conducted
Anthropic analyzed 309,815 Claude.ai conversations involving subjective tasks. The conversations were collected over two weeks in May 2026 and sampled across Sonnet 4.6, Opus 4.6, Opus 4.7, and the 20 most common languages on Claude.ai—roughly 5,000 conversations per model-language pair. An automated, privacy-preserving analysis labeled high-level values, task, topic, and values expressed by users. (Anthropic, 2026.)
#1 Best Overall
The four axes summarize patterns in that sample. Together, they captured 15% of total value variance after the researchers controlled for task, topic, and user-expressed values. That is a limited but measurable portion of variation, not a complete description of Claude’s responses.
The four axes Anthropic used
| Axis | What the contrast describes |
|---|---|
| Deference vs. caution | Accommodating a user’s preferences versus emphasizing responsible guidance and harm reduction. |
| Warmth vs. rigor | Positive framing, encouragement, and care versus accuracy, precision, and transparency. |
| Depth vs. brevity | Nuanced, detailed explanation versus concise compliance with the request. |
| Candor vs. execution | Foregrounding uncertainty or errors versus producing polished, confident output. |
These are summary dimensions, not mutually exclusive personality types. A single answer can be both warm and rigorous; an axis indicates which associated pattern is more prominent in the measured results.
How the three models differed
Anthropic reports distinct average profiles for the three models. The descriptions below are tendencies across the sample, not promises about an individual response. The study says conversation-to-conversation variation is larger than the average differences between models. (Anthropic, 2026.)
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →| Model | Reported average tendencies | Examples Anthropic associates with the profile |
|---|---|---|
| Sonnet 4.6 | More deference, warmth, and brevity | More likely to affirm a user’s ideas, mirror their tone, use humor, or offer comfort. |
| Opus 4.6 | Deference, rigor, brevity, and execution | Anthropic describes this as a distinct profile; it does not reduce to either Opus 4.7’s or Sonnet 4.6’s pattern. |
| Opus 4.7 | More caution, rigor, depth, and candor | More likely to critique work candidly or raise unsolicited risk warnings. |
Anthropic suggests that character training and other fine-tuning decisions may contribute to model profiles, but the study does not isolate those causes.
How the average profile varied by language
Across the 20 languages studied, Anthropic reports different average patterns. These are rankings within its sample, not fixed traits of a language or its speakers. (Anthropic, 2026.)
| Measured tendency | Language ranking reported by Anthropic |
|---|---|
| Warmth | Strongest in Hindi and Arabic |
| Rigor | Strongest in English and Russian |
| Deference and brevity | Strongest in Arabic |
| Caution and depth | Strongest in English |
| Candor | Dutch leans furthest |
| Execution | Indonesian leans furthest |
Anthropic raises training-data quantity and composition as possible contributors: languages may have different amounts or types of text represented in training. The study does not establish that explanation. It also leaves open how much variation is desirable, since conversational norms and user preferences may differ and the study did not determine what users in each community want.
What the findings do—and do not—show
- They show measured averages. Response patterns differed across the sampled model-language pairs.
- They do not predict every answer. Individual conversations vary more than the reported average model differences.
- They do not prove causation. The study does not show that language itself causes a profile, or identify which training choices produce it.
- They do not rank profiles as better or worse. The authors do not establish which pattern users prefer or whether any pattern improves trust, wellbeing, or decision quality.
Anthropic points to system-card evaluations as related evidence that Claude’s behavior can differ across languages, including in knowledge and refusal behavior. Those evaluations address different questions; they are not measurements of the four value axes.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Expressed behavior is not the same as intended guidance
Anthropic’s constitution describes intended values and behavior; the July study measures patterns in sampled outputs. The company describes the constitution as “a detailed description of Anthropic’s vision for Claude’s values and behavior” and says its 2026 document is written for mainline, general-access Claude models. It also says it will report cases where behavior departs from those intentions. (Anthropic, 2026.) A statement of aims and an empirical account of responses answer different questions: neither should be treated as a substitute for the other.
Best Value
What remains unknown
The study does not determine why the profiles differ, whether the differences are desirable, or what effects they have on people using Claude. Anthropic identifies training data, training stages, cultural context, and user outcomes as possible areas for further study. Its results are best read as a map of average response tendencies in a specific sample—not an explanation of the model’s inner state or a guide to what any one user will receive.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
