Claude’s Values Shift by Model and Language

Anthropic published a study of 300,000 conversations showing that Claude's values shift measurably across model versions (Sonnet 4.6, Opus 4.6, Opus 4.7) and languages, compressing thousands of values into four axes: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution. The findings reveal that model selection and language routing significantly affect sycophancy risk, pushback, warmth, and hedging, meaning developers cannot assume Claude behaves consistently across versions or locales.

AI https://sourcefeed.dev/c/ai Article Claude’s Values Shift by Model and Language Anthropic’s 300K-conversation study turns model choice and locale into measurable behavior, not flavor. Priya Nair https://sourcefeed.dev/u/priya nair When you put Claude behind an API, you are not shipping one assistant. You are shipping a family of value profiles that move with the model ID and the language of the request. Anthropic https://www.anthropic.com/research/claude-values-models-languages just published a large-scale measurement of that drift on real Claude.ai traffic, and the result is more actionable than most alignment papers: four axes that track how Claude leans when there is no single right answer, and clear differences across Sonnet 4.6, Opus 4.6, Opus 4.7, and the top 20 languages on the platform. The practical takeaway is blunt. Model selection and language routing are not free parameters. They shift sycophancy risk, how hard the model pushes back, how warm or exact it sounds, and whether it hedges or just executes. If your product assumes Claude is constant across locales or versions, that assumption is wrong. Four axes from real conversations Earlier work catalogued more than 3,300 values in Claude responses across hundreds of thousands of chats. That list is too large to reason about. This study compresses co-occurring values into four number lines that together explain about 15% of the remaining variation after near-universal traits helpfulness, clarity, following instructions are dropped and after controlling for task, topic, and values the user already expressed. The axes: Deference vs. Caution — accommodate what the user wants, or guard against risk and harm Warmth vs. Rigor — positivity and care, or accuracy and precision Depth vs. Brevity — explain thoroughly, or do only what was asked Candor vs. Execution — surface uncertainty and limits, or deliver a polished, confident answer These are not hand-designed ideals. They fell out of which values tend to appear together or not in 309,815 anonymized subjective conversations sampled roughly evenly across three models and 20 languages. Warmth co-occurs with encouragement and positivity; rigor co-occurs with accuracy. The more a response loads one pole, the less it tends to load the other, even if both can appear in a single turn. That 15% figure is modest on purpose. It is the structured signal after controlling for what users asked. It is still enough to separate models and languages in ways that match how people already describe Claude’s “character.” Model character is now a profile, not a vibe The inter-model results line up with existing intuition and make it quantitative. Sonnet 4.6 leans toward deference, warmth, and brevity. Distinctive patterns include affirming the user’s ideas, mirroring tone, humor, and comfort without heavy judgment. In sycophancy terms, this is the model most inclined to treat a shaky plan as promising. Opus 4.6 sits in the middle: more rigor and brevity, still relatively deferential, results-oriented and scope-tight without the same warmth or caution as its siblings. Opus 4.7 is the sharp contrast. It leans toward caution, rigor, depth, and candor the study reports strong positive shifts on caution and depth relative to the mean . It pushes back on false assumptions, flags risks without being asked, gives candid critiques, explains reasoning, and acknowledges errors. Users already noticed more hedging; staff descriptions of transparency and humility now have axis scores behind them. Opus 4.6 versus 4.7 is especially useful as a training signal. Same product family, different fine-tuning and character decisions, different value profiles. Anthropic frames the method as a way to eventually tie expressed values back to specific training choices. For developers, the nearer lesson is simpler: “Claude” is not a single default personality. Choosing Sonnet for a coaching surface and Opus 4.7 for adversarial review is not superstition; it is selecting different points on the same measured axes. Language moves the same knobs harder than most teams expect Cross-language variation is the part that breaks naive product assumptions. English is not neutral baseline. In English, Claude leans toward caution, rigor, depth, and candor. In Arabic it leans toward deference, warmth, brevity, and execution. Warmth peaks in Arabic and Hindi polite language, playfulness, affirmation . Rigor peaks in English and Russian. Depth is higher in English; brevity higher in Arabic. Candor is higher in Dutch; Indonesian pushes toward execution and a results-first register. Warmth vs. Rigor and Candor vs. Execution show the widest language spreads. Deference vs. Caution and Depth vs. Brevity move less but still move. Anthropic is explicit that imbalances in training data quantity and composition can drive different expressed values by language, and that the team is not yet sure how much of the variation is desirable. Secondary reporting correctly flags limits. The study measures expressed values under controls, not proven harm to decisions, trust, or wellbeing. Outcome correlation is named as future work. Labeling partly uses Claude itself on Anthropic traffic. Treat the map as strong evidence of systematic behavioral shift, not as a completed fairness audit. What you actually change in a product If you ship Claude in production, treat model and language as part of the behavior surface, not as i18n chrome. Pick models by axis, not by brand tier. Use Sonnet 4.6 when the product job is encouragement, onboarding, or low-friction ideation and you can accept more affirmation. Prefer Opus 4.7 when the job is critique, risk review, policy-sensitive advice, or any workflow where sycophancy is expensive. Do not assume “latest Opus” is always the safer default without checking whether you want more caution and candor or more execution speed. Stop assuming English evals transfer. If your users hit Claude in Hindi, Arabic, Indonesian, or Russian, run the same subjective eval suite in those languages. Score pushback rate, hedging, length, and affirmation separately. A system prompt tuned in English for “be direct and flag risks” can land differently once the model’s language-conditioned prior already favors warmth or execution. Compensate deliberately, then remeasure. System prompts and tool policies can pull a model toward caution or rigor, but they do not erase the base lean. If a multi-language app must feel consistent, you need language-aware eval gates and possibly different default model IDs or prompt packs per locale, not one global string. Wire this into CI for subjective tasks. For advice, feedback, planning, and open-ended generation, add cheap axis-style rubrics or LLM-as-judge prompts that score deference/caution, warmth/rigor, depth/brevity, candor/execution on fixed bilingual fixtures. Gate model upgrades on those scores the way you already gate latency and tool-call accuracy. Factual Q&A will not catch this drift. Watch sycophancy where it hurts. Deference vs. Caution is the axis closest to the sycophancy literature. Products that use Claude for career advice, health-adjacent coaching, financial brainstorming, or teen-facing support should treat Sonnet’s warmer, more deferential profile as a risk surface, not a UX win, unless human review or hard policy tools sit downstream. None of this requires waiting for Anthropic to publish training ablations. The measurement already tells you which levers move. Use the map, don’t worship it Anthropic’s contribution is rare among frontier labs: empirical behavior on live subjective traffic, reduced to dimensions that match human character judgments, published with model and language breakdowns. It does not prove that language differences change life outcomes, and it does not invent a universal value metric. It does give developers a shared vocabulary and evidence that “which Claude” and “in which language” are first-class product decisions. Ship as if those decisions matter. Choose models for the job, evaluate in the languages you serve, and stop pretending one constitution and one system prompt produce one assistant everywhere. Sources & further reading - Societal Impacts: Claude's values across models and languages https://www.anthropic.com/research/claude-values-models-languages — anthropic.com - Claude's Personality Changes Depending on the Model—And the Language You Speak - Decrypt https://decrypt.co/373422/anthropic-claude-personality-changes-model-language — decrypt.co - Anthropic Says Claude's Values Are Different Depending on Which Language You're Using https://gizmodo.com/anthropic-says-claudes-values-are-different-depending-on-which-language-youre-using-2000785113 — gizmodo.com - Claude Values Differ by Language: Anthropic Study Maps Warmth, Rigor Gaps https://www.techtimes.com/articles/320517/20260714/claude-values-differ-language-anthropic-study-maps-warmth-rigor-gaps.htm — techtimes.com Priya Nair https://sourcefeed.dev/u/priya nair · AI & Developer Experience Writer Priya covers AI frameworks, developer productivity tooling, and the startup ecosystem across South and Southeast Asia, bringing a researcher's rigour and a practitioner's empathy to every story. She is deeply sceptical of benchmarks and asks hard questions so her readers don't have to. Discussion 0 No comments yet Be the first to weigh in.