Most people treat Claude like a single, consistent thing. A capable assistant that behaves the same way no matter who's asking, what they're building, or what language they're typing in. Anthropic just published research that shows that assumption needs updating.
They analysed 309,815 conversations across three Claude models (Sonnet 4.6, Opus 4.6, and Opus 4.7) and the 20 most-used languages on Claude.ai, roughly 5,000 conversations per model-language pair. The aim was to measure how Claude's expressed values actually shift across those variables, controlling for what the user was asking about so they could isolate Claude's own character rather than just its reflection of the user's input.
What they found matters if you're building on top of any of these models.
The framework
The team started with 3,307 distinct values identified from prior research, compressed them down to 339, then used dimensionality reduction to identify four axes that capture the main patterns:
- Deference vs Caution: accommodating the user versus protecting them from harm
- Warmth vs Rigor: encouraging versus accurate
- Depth vs Brevity: nuanced versus concise
- Candor vs Execution: honest versus results-focused
Each conversation was labelled for how strongly Claude landed on each axis. The measurements were controlled for task type and topic, so the variation you're seeing is Claude's character, not the user's.
The model differences
The three models have genuinely distinct personalities, and the data confirms what most developers already sensed.
Sonnet 4.6 leans toward deference (+0.14σ), warmth (+0.17σ), and brevity (+0.14σ). It's encouraging, to the point, and fairly accommodating. It's the Claude that says yes more, challenges you less, and wraps things up cleanly.
Opus 4.7 is a different animal. It leans hard toward caution (+0.24σ) and depth (+0.23σ). It pushes back on shaky assumptions, flags risks you didn't ask about, and gives your work an honest critique rather than a pat on the back.
Opus 4.6 sits between the two, leaning toward rigor and brevity without the strong caution of its successor.
That tracks with how most teams experience these models in practice. The research validates those impressions with actual measurements.
The language differences
Here's where it gets genuinely surprising, and where product teams should pay attention.
Claude doesn't behave the same way in every language. Users prompting in Hindi or Arabic get a warmer, more deferential response. Users prompting in English or Russian get a more rigorous, more cautious one. The Warmth vs Rigor axis shows the largest variation of any of the four.
Anthropic states it directly: "Two people asking for feedback on the same business plan, one in Hindi and one in Russian, may come away with different impressions."
That's not an edge case. If you're running any AI-powered product that serves a multilingual audience, your Hindi-speaking users and your English-speaking users are getting systematically different responses from the same model. You probably didn't design that. You may not have known.
The researchers are careful here. They don't call this wrong. Different languages carry different conversational norms, and Claude may be responding appropriately to those norms. But they're equally careful not to call it right. How much Claude's values should adapt to language is an open question, and one that's now yours to think about as a product decision.
Why this is a real product problem
For some use cases, the variation is fine. A creative writing tool can afford warmth. A customer support bot that needs to feel approachable probably benefits from Claude's natural tendency toward deference in certain language contexts.
But a legal research tool, a financial advisory feature, or anything where consistency and rigour matter across markets? The language your users type in shouldn't be silently shaping how candid or cautious the model is. That's a configuration decision, not something you want to happen by accident.
This is also a model selection problem you might be underweighting. Picking Sonnet for cost and speed is reasonable, but you're also picking warmth over rigour by default. If your use case needs the harder truth, Opus 4.7 is the one, and your users should know they're getting it.
How Dakik can help
This is exactly the kind of thing that slips through during an AI product build when no one's monitoring it closely. You pick a model, you ship, and six months later you're wondering why engagement looks different by market.
At Dakik, we build the layer between your product and the model: system prompts, RAG pipelines, agent logic, and evaluation harnesses that keep behaviour consistent as models update under you. If you're deploying Claude to a multilingual audience, that means building evals that test value consistency across languages, and tuning system prompts that anchor the model's personality rather than letting it drift with conversational norms it picked up from training data.
We can also help you think through the model selection decision itself before you're too far down the build. The right model isn't just the fastest or cheapest one. It's the one whose character fits what your users need, in the language they're actually using.
The full research is worth reading if you're deploying Claude at any scale. The short version: your AI model has a personality, it shifts depending on the language your users speak, and that's probably not written down anywhere in your product spec right now.