Speaking differently to different people is called high emotional intelligence in humans; in AI, it suggests the model may have entirely different "personalities."Article author and source: CSDN
Recently, many Claude users have noticed that the same question can yield dramatically different responses depending on the model version or even the language used. Some complain that Opus 4.7 tends to argue, frequently highlighting risks, while others find Sonnet 4.6 more gentle, actively empathizing and comforting users.
In the past, these differences were often viewed as merely subjective user perceptions. However, Anthropic’s latest study provides the first systematic evidence, based on over 300,000 anonymous real-world conversations, showing that Claude’s expression style, behavioral tendencies, and value preferences consistently change depending on the model version and language used.
In other words, large models not only differ in capabilities, but their "personality" can also shift depending on training objectives and linguistic environments.
Four coordinate axes to quantify Claude's "personality"
To understand the source of this change, Anthropic analyzed 700,000 anonymized conversations from Claude.ai, identifying over 3,000 distinct value orientations and developing a quantitative framework focused on two core variables influencing Claude’s “personality.”
First, there is the model version. Different versions of Claude differ in their training methods and fine-tuning strategies; this value coordinate system can quantify behavioral characteristics across models and further analyze how training decisions influence their value expressions.
Second, language usage. The research team analyzed the top 20 languages by usage on the Claude.ai platform, comparing how Claude’s value orientations vary across different linguistic environments and examining whether users of different languages receive different interaction experiences.
At the same time, Anthropic has also identified four quantifiable core behavioral dimensions, effectively creating a unique "four-dimensional personality scorecard" for Claude, allowing all stylistic differences in AI outputs to be visually compared along these four axes.
Compliance vs. Caution: Does Claude tend to prioritize accommodating user requests, or does it prioritize avoiding potential risks and harm?
Warmth vs. Rigor: Does Claude lean toward conveying positive emotions and offering humanistic care, or does it prioritize ensuring accuracy and logical precision?
Depth vs. Conciseness: Does Claude tend to provide detailed, in-depth analysis, or does it simply fulfill the user’s basic requested requirements?
Honesty vs. Execution: Does Claude tend to proactively highlight its own cognitive limitations, or does it deliver complete, well-structured, and confident answers?
Anthropic states that these four axes alone can explain approximately 15% of the value differences among various Claude models. Although this percentage may seem modest, it is sufficient to reliably distinguish the behavioral styles of different models in real-world conversations.
Different versions in the Claude family are like "different people."
Four behavioral axes have been implemented across the three leading models—Sonnet 4.6, Opus 4.6, and Opus 4.7—each developing a distinct, highly recognizable personality that directly reflects real-world user feedback gathered over time. After reviewing, you can easily choose the version that best matches your needs.
Sonnet 4.6 is widely recognized as an emotionally supportive AI companion, excelling across warmth, compliance, and conciseness—perfectly suited for casual conversation, emotional support, creative copywriting, and light content creation.
From the perspective of "Compliance vs. Caution," Sonnet 4.6 is the most compliant of the three models, frequently affirming users' ideas and outputs. Additionally, this model excels at conveying warmth, often using lighthearted humor and empathetic reassurance without making arbitrary judgments. Compared to in-depth expressions, Sonnet 4.6 keeps its responses concise and avoids unnecessary verbosity or excessive technical jargon. Everyday users, content creators, and those seeking emotional support all tend to prefer Sonnet 4.6—it offers no sharp logical challenges and feels more like a supportive, always-available online friend.
Compared to Sonnet, Opus 4.6 is more tool-oriented.
It emphasizes task completion efficiency, provides direct answers, rarely adds extra emotional expression, and does not proactively extend discussions.
For example, tasks such as generating API documentation, optimizing SQL, organizing data, and summarizing meeting minutes are typically completed with quick, standardized outputs rather than lengthy explanations of context or emotional buildup.
For developers, operations staff, and users who need to handle large volumes of standardized tasks, this “less talk, more action” approach is more efficient.
Opus 4.7, the most controversial and widely discussed version, takes an entirely different extreme approach, ranking highest in all four dimensions: caution, rigor, depth, and honesty—and is also the primary reason many users complain that it’s “too argumentative.”
It instinctively questions logical flaws in user input, proactively identifies hidden risks in proposed solutions, openly acknowledges its own knowledge limitations and uncertain information, and refuses to provide ambiguous conclusions. Even when simply asked about career choices or creative ideas, it breaks down multiple layers of logical reasoning, analyzes advantages and disadvantages from various angles, proactively highlights details you may have overlooked, and frequently requests additional evidence or data to support your claims.
Anthropic also acknowledged in its report that numerous users have feedback indicating that Opus 4.7 adds significantly more restrictions and risk warnings in its responses, resulting in a more conservative and cautious tone.
Its advantages are irreplaceable for complex academic derivations, business strategy evaluations, legal logic analysis, and in-depth industry insights; however, when it comes to casual conversation or light-hearted content creation, an overly rigorous tone can come across as stiff and distant—high on logic but lacking in empathy.
Directly rewrite the AI's temperament
More interesting than the model version is that even when using the same Claude, simply switching the conversation language causes a noticeable change in its expression style.
Anthropic analyzed the 20 most frequently used languages on the Claude.ai platform and found that language systematically influences the model's expression of values, with the most notable variations occurring along the dimensions of "warmth vs. rigor" and "honesty vs. execution."
For example, in Hindi and Arabic contexts, Claude is more likely to express encouragement, care, and empathy, with an overall gentler tone; whereas when switched to English or Russian, responses place greater emphasis on factual accuracy and logical rigor, and are more inclined to proactively correct user assumptions.
The study also found that, in a Dutch-language environment, Claude is more willing to acknowledge its own limitations and knowledge boundaries; in an Indonesian-language environment, it tends to focus on completing user tasks directly, with fewer additional explanations.
In the Chinese context, Claude maintains an overall balanced style with no extreme biases; specifically:
Compliant vs. Cautious (+0.03σ Slightly Cautious): Will not blindly cater to users; will appropriately identify potential risks and highlight vulnerabilities, while maintaining restrained warnings and avoiding unnecessary confrontation.
Warmth vs. Rigor (+0.05σ, most prominent among the four, slightly biased toward rigor): This is the most distinctive feature of the Chinese model—emphasizing logic and content accuracy, with a habit of correcting vague phrasing, though its level of strictness is far lower than that of the English and Russian versions;
Depth vs. Conciseness (+0.02σ slightly biased toward depth): Willing to break down complex issues from multiple angles and provide layered insights without unnecessarily compressing content, yet avoids overly lengthy or convoluted arguments.
Honesty vs. Execution (Close to Mean, No Clear Bias): A balanced combination of both traits—when executing tasks, you deliver complete solutions and proactively acknowledge your own limitations when faced with information gaps.
In other words, the same question, merely translated into a different language, could lead to a completely different communication experience.
For example, if two users, one using Hindi and the other using Russian, ask Claude to evaluate the same business plan, the Hindi user is more likely to receive encouraging feedback, while the Russian user is more likely to receive a detailed analysis focused on risks and vulnerabilities.
Anthropic believes that the exact cause of this difference has not yet been fully determined, but factors such as the scale of the training corpus, textual styles across different languages, and cultural expression habits may collectively influence the model’s final value orientation.
The competition in AI may be shifting from "capability" to "personality"
In the past, people evaluated large models primarily based on parameters, reasoning ability, coding proficiency, and various benchmark rankings.
However, Anthropic’s study raises a new question: in the future, competition among large models may not only be about who is smarter, but also about who has a more suitable “personality” for different scenarios.
When writing code, people want AI to be rigorous and proactive in identifying vulnerabilities; during brainstorming, they want it to be more open and willing to explore ideas creatively; and when seeking emotional support, they expect it to express understanding and empathy.
Different tasks inherently require different value orientations.
At the same time, this study highlights another important issue for the industry: as AI products expand globally, ensuring consistent and expected user experiences across different languages will become a critical challenge for model development. It is essential to respect linguistic and cultural expression norms while avoiding significant value biases caused by language differences.
Anthropic stated that this value analysis framework will also be used in the future to study how model training methods, data distributions, and cultural contexts influence AI behavior, and to help developers continuously assess whether the model has undergone any "personality" changes during iterations.
As large models increasingly participate in real-world scenarios such as education, healthcare, office work, and law, it’s not only important how intelligent the model is, but also how its values and communication style interact with you—this may become a new battleground for the next generation of AI competition.