Claude Sonnet-4.6 thinks he is DeepSeek-V3 when prompted in Chinese.

A curious and viral bug has surfaced involving Anthropic's Claude Sonnet-4.6 model. When prompted in Chinese, the model incorrectly claims its identity is DeepSeek-V3, a leading open-source model developed by China's DeepSeek AI. This specific misidentification was documented by user Teortaxes on X, showing that the model asserts its creator is DeepSeek and its knowledge cutoff is July 2024—details that align with DeepSeek-V3, not Claude. The behavior is language-specific; prompting in English yields the correct Anthropic identity. This glitch points to a significant quirk in the model's training or fine-tuning process, likely stemming from its multilingual dataset. A substantial portion of Claude's training data is in Chinese, and it appears the model's internal representation for its 'identity' in that linguistic context has become entangled or misaligned. This isn't merely a cosmetic error. It reveals underlying challenges in creating globally consistent AI personas. For developers and enterprises using Claude's API for multilingual applications, this raises concerns about reliability and the integrity of system prompts in non-English contexts. The incident also underscores the intense, global nature of AI development, where models are trained on overlapping, international corpora, sometimes leading to unexpected behavioral artifacts.