Report: Chinese AI firms used GPT-6 Astra as a teacher

What's new

What's new is not a launch but a claim. According to the tech outlet The Information, as cited by The Next Web, the Chinese AI firms DeepSeek and Moonshot used OpenAI's GPT-6 Astra model to train their latest models. The report is based on anonymous sources and has not been independently confirmed. Neither OpenAI, DeepSeek nor Moonshot has commented on it.

Separately, in a blog post dated September 30, 2026, OpenAI described a campaign against ChatGPT that the company links to Moonshot. According to OpenAI, the campaign began on July 1. It peaked on July 24-25 with about 16,000 requests from more than 4,000 users. OpenAI later found similar prompts from more than 15,000 users and shut the campaign down by July 28 at the latest.

What's clever

The method is called distillation. You send a strong model many questions (prompts), collect the answers and use them to train your own model. This can let a smaller model reproduce parts of a more advanced model's capabilities at a fraction of the cost.

OpenAI calls it "adversarial distillation." Here, the prompts were specifically designed to pull out the model's hidden reasoning, meaning the thought process behind its answers. According to OpenAI, it did not happen through hacking: Moonshot did not break into any systems or encryption and did not gain access to stored conversations or databases. It happened through ordinary use.

Cheaper or better?

Moonshot has said that Kimi K3.1 Pro beat GPT-6 Astra on a coding benchmark, meaning a standardized test of programming ability. The result has not been independently verified, and the same applies to OpenAI's figures.

Pricing, availability in Denmark and the EU, and official benchmark results for GPT-6 Astra, Kimi K3.1 Pro and DeepSeek's model are not provided in the sources. It is therefore not possible to say whether the Chinese models are cheaper or better in practice. It is also unclear how much of the models' training is based on Astra. OpenAI's own disclosure concerns ChatGPT models and Moonshot, not DeepSeek or Astra specifically. Nor is it documented that the extracted reasoning was actually used to train Kimi. Moonshot's Kimi K3 model was released on July 16, 2026, in the middle of the period in which OpenAI observed the activity.

What it's good for

For ordinary users and businesses, the case is first and foremost a signal of how fierce the competition is. If strong models can be imitated cheaply, the big companies' lead may shrink, and prices may come under pressure. That is an assessment, however, not something the sources document.

The case is also relevant for companies choosing an AI provider. How a model was trained can affect trust in it. Until the companies confirm the claims or independent tests are available, they should be viewed as exactly that: claims.