Chinese artificial intelligence developers DeepSeek and Moonshot AI relied heavily on outputs from OpenAI flagship model GPT 6 Astra to train their latest software. A report from The Information reveals that both developers used model distillation methods to build the Kimi K3.1 Pro and the upgraded V4.1 Flash 1017. The findings point to escalating friction between American research labs and Chinese groups working to lower training costs.
Distillation involves using text generated by an established frontier model to train a newer system. Leading American labs including OpenAI, Anthropic, and Google view this practice as unauthorized capability cloning that allows competitors to replicate advanced reasoning at a fraction of the cost. The method has turned into an ongoing flashpoint in the broader technology contest between the US and China.
OpenAI stated in a public release that security teams traced a wave of coordinated extraction activity back to individuals associated with Moonshot AI. Anthropic also took defensive steps to stop rival developers from copying capabilities out of Claude. Those new security measures cut extracted data traffic by 90% to 99%, marking an aggressive push by top labs to protect proprietary model outputs from unauthorized reuse.
