Even Liang Wenfeng ultimately gave in. DeepSeek’s updated V4.1-Flash model, released on October 17, extensively “distilled” OpenAI’s latest GPT-6 Astra during training. Kimi’s new model from Moonshot AI was also named in this context. Distillation, in simple terms, means using answers from stronger models as teaching material to train your own model—faster progress at lower cost. The results are clear: in a test measuring an AI’s ability to independently perform tasks online, DeepSeek’s new model scored 75.0, while Astra scored 77.9. Yet, the cost per query for DeepSeek is just $0.22, compared to Astra’s $10.77—a nearly 49-fold difference. Previously, Chinese labs primarily distilled Claude’s outputs. But as Anthropic has tightened its defenses, attention has shifted to Astra. Liang Wenfeng originally didn’t intend to rely on distillation for V4. But seeing rivals like Zhipu’s GLM surge ahead using this method, he eventually succumbed to competitive pressure. When DeepSeek first gained attention, it was precisely because it was “cheap and self-trained.” Now, it’s still cheap—but the “self-trained” part is no longer as certain.