Chinese AI Companies Are Using OpenAI’s GPT-6 to Train Their Own Models

Chinese AI Companies Are Using OpenAI’s GPT-6 to Train Their Own Models

There is an interesting development in the global AI race: some of the companies competing with OpenAI may also be learning from its most powerful model.

According to a report by The Information, two major Chinese AI developers, DeepSeek and Moonshot AI, relied heavily on outputs from OpenAI’s GPT-6 Astra while developing their latest models. The report suggests that Astra’s responses became a valuable training resource for these companies, helping them improve their own AI systems.

It is a striking twist in an industry where companies spend billions developing increasingly powerful models. But rather than relying entirely on their own systems, some developers are reportedly using the answers generated by competitors to help train theirs.

So, what exactly happened, and how does it work?

Meet the AI Models at the Centre of the Story

The two Chinese companies named in the report are not newcomers to the AI industry.

DeepSeek gained international attention with models that challenged the assumption that only companies with enormous computing budgets could build competitive AI systems.

Moonshot AI, meanwhile, is behind the Kimi family of AI models, which has gained attention for its reasoning and other advanced capabilities.

Now, both companies are reportedly turning to OpenAI’s GPT-6 Astra as part of their model-development process.

According to the report, Moonshot used Astra during the development of Kimi K3.1 Pro, while DeepSeek reportedly benefited from the same approach when improving its V4.1-Flash model.

The report, published on October 28, 2026, cited people familiar with the companies’ model-development efforts. Reuters also reported on the findings.

Neither company publicly confirmed the specific allegations in the report. 

What Does It Mean to Train an AI Using Another AI?

The technique behind this development is called AI distillation.

The name might sound complicated, but the idea is relatively straightforward.

Think of a student learning from an experienced teacher. The teacher solves difficult problems, explains concepts, and demonstrates how to approach different situations. The student studies those examples and uses them to improve their own performance.

AI distillation follows a similar principle.

A developer gives a more capable AI model different questions and collects its responses. Those responses can then become training examples for another model, helping it learn patterns, improve its answers, and perform better on certain tasks.

For example, a developer could ask a powerful AI system to solve coding problems, explain complex subjects, or work through reasoning tasks. The resulting answers can be used to train another model to produce better responses.

This does not automatically mean the second model becomes identical to the first. It means the developer can use the first model's outputs as learning material.

It is also important to understand that distillation is not automatically wrong. AI companies use the technique themselves to create smaller, faster, and less expensive versions of their own models.

The controversy begins when a company uses another developer's protected model to extract its capabilities without permission, particularly when the model's terms of service prohibit that activity. 

Why GPT-6 Astra Became an Attractive Target

If a company has already invested heavily in building a powerful AI model, why would another company want to learn from it?

The answer is simple: developing advanced AI from scratch is expensive, time-consuming, and technically demanding.

Training a frontier model requires enormous computing resources, large amounts of data, and specialized research teams. Even after all that investment, there is no guarantee that the final product will perform as expected. Learning from an existing powerful model can help developers improve certain capabilities without having to discover every useful technique independently.

In this case, GPT-6 Astra reportedly became a major target after Anthropic increased its efforts to prevent unauthorized extraction of capabilities from its Claude models.

The Information reported that Chinese AI developers had previously relied heavily on Claude for distillation. As Anthropic strengthened its protections, attention reportedly shifted towards OpenAI's Astra.

OpenAI says it is monitoring these activities and taking steps to identify and disrupt attempts to extract its models' capabilities. 

OpenAI Is Not Taking the Situation Lightly

For OpenAI, this is more than a question of who uses its chatbot.

The company invests heavily in developing its models. Their reasoning abilities, coding performance, and other capabilities are part of what gives them commercial value.

If competitors can extract those capabilities at scale and use them to strengthen their own products, it raises questions about intellectual property, fair competition and the cost of building advanced AI.

OpenAI has reportedly responded by restricting suspicious accounts, strengthening sign-up and infrastructure controls, and expanding its monitoring systems.

The company has also previously disclosed a coordinated model-distillation campaign involving individuals associated with Moonshot. It said the activity was identified and mitigated in July.

Anthropic has taken similar action against what it describes as large-scale, unauthorized distillation campaigns targeting Claude.

These developments show that protecting AI models is becoming an increasingly important part of the competition between leading AI companies.