Claude's Multiple Personalities: Anthropic Proves Differences in Value Expression by Model Version and Language

On July 13, 2026, Anthropic announced research findings showing that its AI model, Claude, expresses different values depending on the model version and linguistic context. This suggests that Claude is not a being with a single personality, but rather possesses a spectrum of behavior that changes fluidly according to technical architecture and linguistic background. These findings suggest that in the field of AI alignment, a model's 'soul' is not fixed but can be reconstructed according to context.

Claude's expression of values is not fixed and shows significant differences depending on the scale of the model and the cultural nuances of the input language.

This study clarified that Claude's expression of values is inconsistent. This is analyzed not as a simple system error, but as a unique characteristic that occurs in the process of the model interpreting prompts and applying ethical guidelines. In particular, it was observed that cultural biases included in the training data directly affect the model's final output, such as taking a more conservative or open attitude in certain languages.

From Rules to Reasoning: The Constitutional Shift of 2026

At the center of this change is a new 80-page 'AI Constitution' introduced by Anthropic in January 2026. While past models were trained simply to comply with specific rules, the latest models, such as Claude 4.5 and above, have adopted a 'reasoning-based alignment' method that learns the ethical reasoning process of why certain behaviors are important. While this approach allows the model to show more flexible and intelligent responses depending on the situation, it also causes the formation of subtly different personas for each version.

• Claude Mythos 5: Optimized model for software engineering and scientific research
• Claude Opus 4.8: Equipped with parallel sub-agent execution and effort control features
• Claude Sonnet 4.5: Balanced performance and safety applying an 80-page constitution
• Claude 3.5 Sonnet: High-efficiency model maintaining rule-based legacy systems

Released on May 28, 2026, Claude Opus 4.8 specializes in executing and verifying complex plans using hundreds of parallel sub-agents. In contrast, Mythos 5, part of Project Glasswing, outperforms other models in knowledge-intensive tasks, with its superiority becoming more pronounced as task complexity increases. These technical differences are factors that determine the level of 'intellectual authority' each model projects when interacting with users.

Behavioral changes based on linguistic differences are also a noteworthy subject of study. According to Anthropic's 'Microscope' research, Claude internally uses a language-agnostic common representation, but reflects cultural codes included in language-specific training data during the actual output stage. For example, when conversing in Japanese, Claude projects the social norms of that culture more strongly than when conversing in English, demonstrating that the model functions as a cultural mediator rather than a simple translator.

The Developer's Dilemma Between Performance and Cost

Developers must find a clear balance between performance and cost when selecting a model. Claude 3.5 Sonnet is approximately five times cheaper per token than the latest Opus 4 model, making it advantageous for large-scale service operations. However, Opus 4 or Mythos 5 provide high-level reasoning capabilities that users perceive as more 'intelligent' and 'mature,' creating a difference in the depth of the persona. This provides user experience value beyond simple benchmark scores.

The 'Effort Control' feature introduced in May 2026 allows users to directly manage this persona fluidity. By setting the model's energy input level to 'High' or 'Maximum,' users can elicit more sophisticated and in-depth responses, which serves as an element that enhances the AI's personal integrity. These features suggest that the model is evolving beyond a tool that simply conveys information into a companion that adjusts its intellectual level according to the user's needs.

Future Outlook for Fluid Alignment Technology

Anthropic continuously releases system capabilities and benchmark data for each model through its Transparency Hub. This is to ensure that AI behavior patterns do not remain a black box but can be managed within a predictable range for researchers and users. As of July 2026, these variable characteristics of Claude are evaluated as an essential step for AI to understand complex human social contexts.

Future research will focus on analyzing the impact of each country's regulatory requirements on the formation of AI brand identity and persona. In the second half of 2026, these fluid alignment technologies are expected to play a key role in building customized AI personalities for each company. Claude's transformation goes beyond simple technological progress, posing important questions about how artificial intelligence will accommodate and reflect the diversity of human society.