Foreign media report that DeepSeek and Kimi distilled Astra during training.

According to The Information, citing multiple sources with direct knowledge, DeepSeek’s AI company released version V4.1-Flash-1017 on October 17, and Moonshot’s recently launched Kimi K3.1 Pro both employed OpenAI’s GPT-6 Astra outputs at scale during training—via knowledge distillation. Notably, DeepSeek’s new model shows improved performance on browser-operation tasks: on Browser Use’s benchmark suite covering 180 real-world web scenarios, its score rose from 70.9 in the prior version to 75.0—approaching Astra’s 77.9. Meanwhile, the average per-task model invocation cost stands at $0.22, versus Astra’s $10.77—a mere ~1/49th of the latter. Additionally, insiders noted that this move…