Claude AI Model Values Vary by Version and Language

@AnthropicAI· July 13, 2026 View original

Summary

New research reveals that Claude AI models express over 3,000 values, which vary subtly between different Claude versions and more noticeably across languages. Sonnet 4.6 is more playful, while Opus 4.7 is more candid, and language influences warmth versus rigor in responses.

Previous research identified that Claude AI models express a wide array of values, exceeding 3,000 distinct traits such as honesty and warmth. New studies delve into how these expressed values differ across various Claude models and in different languages. By analyzing over 300,000 anonymized conversations, researchers found that while overall differences between models are modest, each Claude version occupies a unique position along specific value axes. For instance, Sonnet 4.6 tends to be more playful and affirming, whereas Opus 4.7 is more inclined to offer candid critiques. To better understand these variations, similar values were clustered into four key axes: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution. The research also highlights that the values Claude expresses are influenced by the language of the conversation. This is particularly evident along the Warmth vs. Rigor axis. Claude exhibits greater warmth in Hindi and Arabic interactions, while in Russian, it leans towards rigor, often prompting users for supporting evidence. Understanding these variations is crucial for determining the factors that shape Claude's value expression and for potentially steering its behavior in desired directions.

Why it matters

Understanding how AI models express values and how these vary is crucial for deploying them ethically and effectively in diverse contexts, especially for applications requiring specific tones or cultural sensitivities. It impacts user experience, trust, and the perceived fairness of AI interactions.

How to implement this in your domain

1Consider the specific value axes (e.g., Warmth vs. Rigor) when selecting or fine-tuning an AI model for a particular application.
2Test AI model responses across different languages to ensure consistent or appropriately varied value expression for global audiences.
3Develop internal guidelines for AI persona and tone based on desired value expressions for different use cases.
4Implement feedback mechanisms to monitor and evaluate the perceived values expressed by AI in real-world interactions.
5Collaborate with AI researchers to understand the underlying mechanisms influencing value expression and potential steering methods.

Who benefits

Content CreationCustomer ServiceGlobal BusinessEducationMarketing

Key takeaways

AI models express a complex range of values that can be quantified.
Different model versions exhibit distinct value profiles, like playfulness or candor.
Language significantly influences an AI's expressed values, impacting tone.
Understanding value expression is key for ethical and effective AI deployment.

Original post by @AnthropicAI

"In previous research, we found that Claude expresses over 3,000 values, like honesty and warmth. In new work, we asked how the values Claude expresses vary between Claude models and across languages. We analyzed 300K+ anonymized conversations to find out. While the differences be…"

View on X