DeepSeek and Moonshot train models on OpenAI's GPT-6 Astra responses, The Information reports

According to the publication, three sources spoke about the use of Astra in the creation of Kimi K3.1 Pro, which Moonshot AI released on October 23. Another two reported that responses from this model were used in the DeepSeek V4.1-Flash 1017 update on October 17. Moonshot did not comment on the publication, and DeepSeek did not respond to an inquiry. As 3DNews reports citing Reuters, representatives of OpenAI also did not respond to the agency's questions.

This involves distillation: responses from a more powerful model serve as training material for another system. It remains unknown how the Astra responses were obtained and at which training stages they were applied: The Information describes neither the process nor the composition of the data.

According to the publication, one of the main reasons for switching to Astra was Anthropic's countermeasures. On October 12, the company reported that after strengthening Claude's defenses, the number of extracted hidden reasoning tokens dropped by 90–99%. Earlier, on February 23, Anthropic accused DeepSeek, Moonshot, and MiniMax of mass distillation of Claude: according to its data, these labs conducted over 16 million chats through roughly 24,000 fake accounts.

On September 30, OpenAI revealed a campaign that began in July, part of whose activity the company linked to individuals associated with Moonshot AI. Associated query patterns were detected across more than 15,000 users; by July 28, the campaign was halted, according to OpenAI. On November 20, a new version of OpenAI's terms comes into effect, where the ban on training competing systems explicitly mentions the generation of training data, evaluation, and ranking of responses.