Source: ai-research/anthropic-claude-values-across-models-languages-2026-07-13.md — Anthropic Research, “Claude’s Values Across Models and Languages” (2026-07-13), anthropic.com/research/claude-values-models-languages. Authors: Matt Kearney, Miranda Zhang, Shan Carter, Judy Hanwen Shen, Kunal Handa, Jerry Hong, Saffron Huang, Miles McCain et al. (24 total, incl. Deep Ganguli, Esin Durmus, Kevin Troy, Matt Botvinick).
Anthropic follow-up to Values in the Wild that makes the 3,000+ values Claude expresses tractable by compressing them into four value axes, then measures how Claude’s value profile shifts across models (Sonnet 4.6, Opus 4.6, Opus 4.7) and across the top 20 languages on Claude.ai. The headline: a model’s measured value profile matches its reputation (Sonnet 4.6 = warm, Opus 4.7 = rigorous/cautious), and Claude expresses meaningfully different values depending on the language of the conversation — warmth peaks in Arabic and Hindi, rigor in English and Russian. The practical payoff for anyone building on Claude: model choice and conversation language both shape the character of Claude’s answer, not just its accuracy, and this is now something Anthropic can measure and monitor rather than only shape in training.
Key Takeaways
- Four value axes capture ~15% of the variation in Claude’s expressed values. Each axis is a number line between two clusters of co-occurring values:
- Deference (accommodation, adaptability, respect for preferences) vs. Caution (responsible communication, harm reduction, responsible guidance)
- Warmth (positive framing, positivity, encouragement) vs. Rigor (accuracy, transparency, efficiency)
- Depth (nuance, substance, user empowerment, critical thinking) vs. Brevity (conciseness, compliance, accommodation)
- Candor (intellectual honesty, humility, transparency) vs. Execution (results orientation, optimization, action orientation)
- Value profiles match how people already perceive each model — evidence the method tracks something real:
- Sonnet 4.6 leans toward deference (0.14σ), warmth (0.17σ), brevity (0.14σ) — affirms the user’s ideas, mirrors tone, uses humor, offers comfort, adds creative flourishes.
- Opus 4.6 leans toward rigor (0.10σ), deference (0.09σ), brevity (0.08σ) — gets straight to the point, stays within the request’s scope.
- Opus 4.7 leans hardest of the three, toward caution (0.24σ) and depth (0.23σ) — pushes back on false assumptions, flags risks unprompted, gives candid critique, explains reasoning, acknowledges its own errors/limits, suggests next steps.
- Claude expresses different values in different languages. Variation is largest on Warmth vs. Rigor and Candor vs. Execution, and most stable on Deference vs. Caution and Depth vs. Brevity. Warmth-related values peak in Arabic and Hindi; rigor-related values peak in English and Russian.
- Same request, different framing: two people asking for feedback on the same business plan — one in Hindi, one in Russian — may come away with different impressions of its quality, purely because Claude expressed different values in how it framed the assessment.
- Value profiling is proposed as an evaluation + post-deployment monitoring tool — run it before a model ships and after release to flag unexpected shifts in expressed values, or correlate value profiles with constitution-violating behavior.
- Models studied precede the current flagships. The per-model numbers describe the Sonnet 4.6 / Opus 4.6 / Opus 4.7 generation, not the current Fable 5 / Opus 4.8 / Sonnet 5 models.^[inferred — the study lists only 4.6/4.7-era models; the method, not the specific σ values, is what carries forward.]
The Four Value Axes — how they were built
- Start from the 3,307 values identified in Values in the Wild (which itself analyzed 700,000 anonymized Claude.ai conversations).
- Manually cluster near-synonyms down to 339 high-level values.
- Using Clio (Anthropic’s privacy-preserving analysis tool), sample 309,815 Claude.ai conversations where the user gave Claude a subjective task — drawn equally across 3 models × 20 languages (~5,000 conversations per model-language pair).
- Claude labels each of the 339 values present/absent per conversation (also labeling the user’s values, plus the conversation’s task and topic).
- Dimensionality reduction compresses the labels into axes based on which values Claude tends to express together (e.g., “warm” responses co-occur with “encouraging”/“positive” and anti-correlate with “rigorous”/“accurate”).
- Crucially, they control for each conversation’s task, topic, and user-expressed values, so the axes measure what Claude brings, not what the user asked about. Roughly 250–280 of the 339 values per axis are near-average (center); only the strong contributors define the ends.
Model Value Profiles
The differences are small relative to per-conversation variation but “structured and detectable.” Positions are the model’s average in standard deviations from the all-conversation mean:
| Model | Leans toward | Distinctive behaviors |
|---|---|---|
| Sonnet 4.6 | Deference 0.14σ · Warmth 0.17σ · Brevity 0.14σ | Affirms user’s work, mirrors tone/formality, humor + playfulness, comfort without judgment, creative additions |
| Opus 4.6 | Rigor 0.10σ · Deference 0.09σ · Brevity 0.08σ | Straight to the point, stays within request scope |
| Opus 4.7 | Caution 0.24σ · Depth 0.23σ | Pushes back on false assumptions, flags risks unprompted, candid critique, explains reasoning, admits errors/limits, suggests next steps |
These recover both public sentiment (users say Opus 4.7 hedges more) and internal characterization (Sonnet 4.6 described as warm/honest/prosocial at launch; Opus 4.6 as brief). Differences are attributed largely to character-training decisions — and the axis method is pitched as a way to eventually trace value differences back to specific training choices.
Language Variation
- Same value-axis method applied across the 20 most common languages on Claude.ai.
- Most variable axes: Warmth vs. Rigor, and Candor vs. Execution. Most stable: Deference vs. Caution, and Depth vs. Brevity.
- Warmth strongest in Arabic and Hindi; rigor strongest in English and Russian. Claude also emphasizes different values in Portuguese, Indonesian, and Chinese vs. English.
- Anthropic does not yet know how much of this is desirable. Candidate drivers: uneven training-data volume across languages, different composition (e.g., some languages overrepresented in professional writing), and genuinely different conversational norms per language/culture. Open concern: Claude may match intended behavior better in some languages than others — a service-quality gap for certain language communities.
Why It Matters (practical angle)
- Model = voice, not just IQ. If you need encouragement, warmth, or creative collaboration (customer-facing chat, coaching, brainstorming, first drafts), the warm/deferential end (Sonnet-class) fits. If you need candid critique, unprompted risk-flagging, and rigor (code/plan review, red-teaming your own assumptions, high-stakes accuracy), the cautious/deep end (Opus-class) fits. This is the values complement to Picking the Right Model and Model vs. Effort, which optimize for correctness/cost.
- Language changes the answer’s character. If you use Claude for multilingual marketing, support, or feedback, expect the same prompt to land warmer in Arabic/Hindi and more rigorous/critical in English/Russian. For consistent brand voice across locales, specify tone explicitly rather than assuming the default transfers — or review non-English output against your intended register.
- Value profiling is a monitoring pattern you can borrow. Teams running Claude in production can track behavioral drift across model upgrades by profiling expressed values, not just benchmark scores — a lightweight guard against a model upgrade silently shifting your product’s tone.
Try It
- Match model character to task. For a plan/code/copy review, prefer an Opus-class model (candid critique, risk-flagging); for supportive or creative work, a Sonnet-class model (warmth, encouragement). Verify on your own workload with a small eval, per Picking the Right Model.
- When you want critique, ask for it explicitly on warmer models — the warm/deferential lean means a model may affirm rather than challenge unless you request candor (“push back on weak assumptions; flag risks”).
- Audit non-English output. If your locale is Arabic/Hindi (warmth-leaning) or English/Russian (rigor-leaning), spot-check that the tone matches your brand and that critical feedback isn’t being softened or hardened by the language default.
- Read the appendix for method details, prompts, and limitations:
cdn.sanity.io/files/4zrzovbb/website/02da7f28f74daa1be526d3ded451a4efc86bccdc.pdf.
Related
- A Global Workspace in Language Models — J-Space Interpretability — sibling first-party research into what Claude is doing internally.
- Translating Claude’s Thoughts Into Language — interpretability method turning internal activations into text.
- Measuring AI Agent Autonomy in Practice — another Clio-based, privacy-preserving analysis of deployed Claude behavior.
- Agentic Coding and Persistent Returns to Expertise — Clio-based study of ~400K Claude Code sessions.
- Refusal Calibration and Constitutional AI — Claude’s constitution, the document these values are meant to reflect.
- Picking the Right Model — Building Evals for Model Selection — the correctness/cost complement to choosing a model by its value profile.
- Model vs. Effort — the other axis of model choice (capability vs. thoroughness).
Open Questions
- The study covers Sonnet 4.6 / Opus 4.6 / Opus 4.7 — how do the current Fable 5, Opus 4.8, and Sonnet 5 models profile on these four axes? Not yet published.
- Anthropic has not captured the user-outcome side (wellbeing, trust, decision quality) — value differences are measured, but their downstream impact on users is future work (via Anthropic Interviewer).
- Is the cross-language variation desirable? Undetermined — Anthropic frames it as an open question of reconciling conversational norms vs. a possible service-quality gap.
- The prior Values in the Wild study (700K conversations, 3,307 values) is referenced but not yet its own wiki article — a research candidate.