Foreign media reports that DeepSeek and Kimi are distilling Astra during training

According to The Information and multiple direct sources, the artificial intelligence company DeepSeek updated its V4.1-Flash-1017 on October 17, and the newly released Kimi K3.1 Pro from Dark Side of the Moon both heavily utilized outputs from OpenAI GPT-6 Astra for distillation during training.

The new version of DeepSeek scored higher in browser operation task tests. The 180 real webpage task tests for Browser Use showed its score increased from 70.9 in the old version to 75.0, approaching Astra's 77.9; the cost per model call was $0.22, while Astra was $10.77, nearly a 49-fold difference.

Informed sources also mentioned that previously Claude was the main target for distillation by domestic AI companies, but Anthropic has tightened its defenses in the past two months, making it more difficult to extract the model inference process, leading domestic laboratories to shift their focus to Astra. DeepSeek's founder Liang Wenfeng was initially reluctant to rely on distillation in the development of V4, but after seeing domestic competitors like Zhipu GLM make progress through distillation, he compromised under competitive pressure.