DeepSeek and Kimi copied GPT-6: China is training its AI on OpenAI
The new Kimi K3.1 Pro and DeepSeek V4.1-Flash were trained on GPT-6 Astra outputs. Chinese labs milked Claude the same way until Anthropic shut the tap.
Two of China's best known AI labs, DeepSeek and Moonshot AI, leaned heavily on answers from GPT-6 Astra, OpenAI's most powerful model, while training their latest releases. That is according to US reporters citing people with direct knowledge of the matter.
Three sources said Moonshot used Astra to build Kimi K3.1 Pro, released on October 23. Two more said the same was true of DeepSeek V4.1-Flash 1017, an update shipped on October 17. Moonshot declined to comment and DeepSeek did not respond.
The technique is called distillation: a strong model produces solutions, explanations and reasoning, and a weaker one learns from them. It is a normal, legal method in itself, but here it sidesteps the terms of access to someone else's model. For Chinese teams it saves months of work and piles of compute they are already short on because of limits on US accelerators. And the finished model runs on their own servers with no dependence on OpenAI's API.
Claude used to be the main donor for such schemes. Anthropic tightened its defenses and says the volume of harvested data fell by 90 to 99 percent. The Chinese labs simply moved on to OpenAI. OpenAI itself said on September 30 that it had tied the core of a distillation campaign that started in July to people linked to Moonshot, and on October 21 it announced that routers and API gateways offering its models will need its approval and will have to help catch abusers.
The catch is that labs want to sell access to their best models, yet every answer can become a textbook for a future rival. Ordinary API customers end up paying for it, since every new restriction lands on them too.
Verdict: an AI race where one side builds and the other copies.