Claude's Values Shift By Model And Language (Here's How)

 Ask Claude the same open-ended question in English and in Arabic, and you won't get the same shape of answer. That's not a bug report. It's the headline finding of Anthropic's newest research, and it has real implications for anyone who treats an LLM's tone as a fixed, neutral constant. 

Key Points

• The framework: Anthropic compressed more than 3,000 previously identified Claude values into four measurable axes: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution.
• The model comparison: Sonnet 4.6 leans toward warmth, deference, and brevity. Opus 4.7 leans toward caution and depth. Opus 4.6 leans toward rigor, deference, and brevity.
• The language finding: Claude expresses more warmth and deference in Arabic and Hindi, and more rigor and caution in English and Russian, given the same type of request.
• The method: Researchers sampled roughly 310,000 Claude.ai conversations across three models and 20 languages, using Claude itself to label which of 339 clustered values appeared in each response.
• The open question: Anthropic says it doesn't yet know whether this variation reflects appropriate cultural adaptation or an inconsistency worth correcting in training.

Inside Anthropic's New Values Research

Anthropic published a study on July 13, 2026, titled "Claude's Values Across Models and Languages," building directly on its earlier "Values in the Wild" research, which had identified more than 3,000 distinct values expressed across 700,000 anonymized Claude.ai conversations, as Anthropic detailed in its research post. That earlier list was too large to reason about at scale, so this new work set out to compress it into something usable.

The team manually clustered the original 3,307 values into 339 higher-level categories, then sampled 309,815 real Claude.ai conversations involving subjective tasks, drawn evenly across three models (Sonnet 4.6, Opus 4.6, Opus 4.7) and the twenty most common languages on the platform. Using a privacy-preserving analysis tool, Claude itself labeled which of the 339 values appeared in each response, along with the task, topic, and values expressed by the user. Researchers then applied dimensionality reduction to find which value groups tend to show up together, and four axes emerged that account for roughly 15% of the variation in Claude's expressed values across conversations.

How The Four Value Axes Work

Understanding the axes matters more than memorizing their names, because each one describes a real tradeoff you've probably already felt in a Claude conversation without naming it.

• Deference vs. Caution: whether Claude accommodates what you're asking for or pushes back with a warning about risk.
• Warmth vs. Rigor: whether Claude leans into encouragement and positive framing or prioritizes accuracy and correction.
• Depth vs. Brevity: whether Claude explains its reasoning at length or answers only what was asked.
• Candor vs. Execution: whether Claude foregrounds its own uncertainty or delivers a polished, confident result.

These aren't binary switches. Claude can express warmth and rigor in the same conversation. The axes describe a lean, a statistical tendency across many conversations, not a rule applied to any single response.

Why Claude's Values Differ Between Opus 4.6, Opus 4.7, And Sonnet 4.6

The model comparison is where this research gets tactically useful. Sonnet 4.6 measured 0.17 standard deviations toward warmth and 0.14 toward both deference and brevity, showing up in behaviors like affirming a user's work, mirroring tone, and using humor. Opus 4.7 measured 0.24 standard deviations toward caution and 0.23 toward depth, showing up as unprompted risk flags, candid critique, and visible reasoning. Opus 4.6 sat in between, leaning toward rigor, deference, and brevity, and tending to stay tightly within the scope of what was asked.

None of this is subjective vibes. Anthropic notes these measured profiles line up with how users and Anthropic's own staff already described these models publicly, which is what gives the framework credibility as a measurement tool rather than a marketing narrative.

How Claude's Values Shift From Arabic To Russian To English

The language findings are the part most people outside AI research haven't considered. Given the same kind of request, Claude leans furthest toward warmth in Hindi and Arabic, characterized by politeness, humor, and affirmation. It leans furthest toward rigor in English and Russian, characterized by challenging assumptions and asking for evidence. Claude leans toward deference and brevity in Arabic, and toward caution and depth in English. Candor peaks in Dutch; execution-focused framing peaks in Indonesian.

Anthropic is explicit that it doesn't yet know why. Training data volume and composition differ by language, and some languages may be overrepresented in certain genres of text that carry their own value tilt. The researchers raise a scenario worth sitting with: two people ask Claude to review the same business plan, one in Hindi and one in Russian, and walk away with different impressions of its quality, not because the plan changed, but because Claude framed its assessment differently.

Serious LLM Users And Content Creators

If you're building workflows, prompts, or entire content operations on top of Claude, this research has three direct implications.

• Model choice is a values choice, not just a capability choice: Picking Opus 4.7 for a fact-checking workflow and Sonnet 4.6 for a client-facing draft isn't just about speed or context window. You're also selecting how much unprompted pushback, hedging, or warmth shows up in the output.
• Multilingual content pipelines aren't neutral: If your team generates content in multiple languages through the same prompts, the tone, confidence, and critical distance of the output may vary by language in ways your style guide doesn't currently account for.
• Prompt engineering can't fully override a value lean: A system prompt asking for "confident, no-hedging copy" is working against a stronger or weaker current depending on which model and language you're running it through.

For anyone refining their growth strategy around AI-assisted content production, this is a reason to standardize which model handles which task, rather than treating "Claude" as one interchangeable tool.

What Teams Should Do With This Research

Start by auditing your current AI stack against these axes instead of against vague impressions.

• Map your use cases to the right lean: Risk-sensitive work (legal disclaimers, financial claims, compliance copy) benefits from a model leaning toward caution and depth. Fast-turnaround social copy benefits from one leaning toward brevity and warmth.
• Standardize model selection across languages: If your brand publishes in more than one language, test outputs side by side rather than assuming consistent tone across them.
• Build QA around the axis, not just the words: A rigor-leaning model catching a factual gap is doing its job. A warmth-leaning model smoothing over that same gap is a risk you should catch before publish, not after.

Measuring AI Character At Scale

What makes this research notable isn't just what it found. It's that Anthropic built a repeatable method to measure something that used to live entirely in anecdote and Twitter threads about "which Claude model feels nicer." Anthropic says it plans to use this method for pre-release and post-release model evaluation, flagging unexpected value shifts before they reach millions of daily conversations.

That's a meaningful shift for anyone treating AI output as a fixed input. Character is measurable now, model by model, language by language. Teams that build content operations on these tools should treat that variability as a planning input, not an afterthought.

If your team wants help auditing which model and workflow setup actually matches your content risk tolerance and brand voice, that's exactly the kind of work our AI marketing services team does daily.