Report says Chinese AI labs leaned on OpenAI's answers to train their models

What the report says

Global Banking & Finance Review carried a story under the headline "Chinese AI Firms Use OpenAI's GPT-6 Astra for Model Training: Report." The underlying reporting comes from The Information. Reuters, via TradingView, and the Economic Times also carried it.

According to that reporting, two Chinese AI companies, DeepSeek and Moonshot AI, relied heavily on answers from OpenAI's models to train their newest models. The report cites people with knowledge of the matter. It does not name them in the coverage we saw.

The word report matters here. Neither company's side of the story appears in the material we reviewed, and nothing we saw confirms the claim independently.

How one model can learn from another

A quick definition first. Training is the process of showing an AI model a huge number of examples until it gets good at producing useful answers. The quality and variety of those examples shape how capable the model ends up.

One way to get examples is to ask an existing, strong model many questions and keep its answers. The new model then studies those answers and learns to imitate them. People in the field often call this general approach distillation, a word that suggests boiling a big model's skill down into something smaller or cheaper.

Think of a student who learns by reading a top tutor's worked solutions instead of working through textbooks alone. The student can improve fast, but the result depends heavily on the tutor. That is the idea behind the phrase in the report: that these labs relied heavily on OpenAI's answers.

What we do not know yet

The gaps are large, and it is worth listing them plainly.

• The coverage we saw does not confirm the claim.
• The sources are unnamed people with knowledge of the matter.
• We have no numbers on how much of the training used OpenAI's answers.
• The coverage we saw does not say how the answers were collected or what rules, if any, were involved.

The headline names GPT-6 Astra as the OpenAI model in question. The other coverage we saw refers only to OpenAI's models in general. So the exact model is one more detail to treat carefully until something firmer appears.

Why a general reader should care

Most people choose AI tools by price, speed and how good the answers feel. A claim like this adds another question: where did a model's skill come from? A cheaper model that learned from a pricier one can look like a bargain, but it may also share the strengths and weaknesses of its teacher.

It also bears on the wider race between AI labs. If building a strong model can lean on a rival's outputs, the gap between a leader and a follower may be smaller than spending alone suggests. That is a big "if," because the report is unconfirmed.

For everyday use, the practical lesson is modest. Different tools may agree with each other because they share ancestry, not because they independently checked the facts. Agreement across tools is not proof that an answer is right.

What to watch next

The first thing to look for is a response from the named companies, or from OpenAI. Any of them could confirm, dispute or add detail to the report.

The second is whether other outlets with their own sources report the same thing. One report, even one carried by many outlets, can trace back to a single original source. Wide coverage is not the same as wide confirmation.

Until then, hold the story loosely. It is a credible lead worth following, not a settled fact.