According to The Information, Chinese AI firms DeepSeek and Moonshot AI utilized GPT-6 Astra to train their latest models. The affected systems include the Kimi K3.1 Pro, released by Moonshot last week, and the updated DeepSeek V4.1-Flash, which debuted on October 17.
The companies reportedly employed model distillation, a technique where developers train a smaller system using outputs generated by a more powerful model. This method allows firms to replicate specific capabilities of advanced AI without incurring the massive costs associated with building such models from scratch.
OpenAI previously accused Moonshot of similar practices. On September 30, OpenAI reported that in July, users linked to the Kimi developer attempted to extract hidden reasoning data from its models. The company identified over 15,000 accounts involved in this activity and blocked them by July 28.
Major US developers like OpenAI, Anthropic, and Google are actively working to prevent such data extraction. In early October, Anthropic announced that its new security measures successfully reduced the volume of data harvested from Claude models for distillation purposes by approximately 90 to 99 percent.
Both Chinese companies have reported strong performance metrics for their new models, though these results remain unverified by independent sources. Moonshot claims Kimi K3.1 Pro outperformed GPT-6 Astra in a coding benchmark, while DeepSeek positions its V4.1-Flash as a cost-effective alternative to American AI systems.
