Even Liang Wenfeng couldn't withstand it in the end. DeepSeek's V4.1-Flash model, updated on October 17th, heavily "distilled" OpenAI's latest GPT-6 Astra during training. Kimi's new model from the Dark Side of the Moon was also mentioned. Distillation, simply put, is using the answers of stronger models as teaching materials to feed your own model. It's fast and inexpensive. The results are direct. In a test assessing an AI's ability to perform online tasks, the new DeepSeek version scored 75.0, while Astra scored 77.9. However, the cost per call was only $0.22 for DeepSeek, compared to $10.77 for Astra—a difference of nearly 49 times. Previously, domestic labs mainly distilled Claude. Recently, with Anthropic's stricter controls, everyone has turned their attention to Astra. Liang Wenfeng originally didn't want to rely on this method for V4. Later, seeing competitors like GLM using distillation to advance, he eventually succumbed to competitive pressure. DeepSeek initially gained popularity because of its "cheap and self-taught" approach. Now that the cheapness remains, the latter part of that claim is less certain.