Reported: DeepSeek and Moonshot Used OpenAI's GPT-6 Astra to Train Their Latest Models

Two of China's leading AI companies reportedly used outputs from OpenAI's most powerful model to train their newest systems. That is the account published by Reuters on 29 October 2026, reporting The Information's reporting, which cited people with direct knowledge of the matter. DeepSeek and Moonshot AI, the report says, relied heavily on outputs from OpenAI's GPT-6 Astra model to train their latest models. Neither company, nor OpenAI, had responded to Reuters' requests for comment at the time of publication. Everything in this article that describes what the two companies did is that reported account, not a proven fact — no evidence has been published, no company has confirmed it, and no legal finding exists.

The two models named are both recent. Moonshot AI's Kimi K3.1 Pro, released in the week before the report, and DeepSeek's upgraded V4.1-Flash 1017, launched on 17 October 2026, both benefited from distilling Astra, The Information said, according to Reuters. Distillation means training a new model on the outputs of an existing, more capable model, so the new model reproduces some of its capabilities without the same training from scratch. The technique itself is standard and used legitimately, including by the leading labs. The controversy is about scale, method and permission: whether outputs were collected systematically, whether access restrictions were bypassed, and what the provider's terms allowed.

What was reported, and what was not

Reuters' report is short, and its limits are visible. It says DeepSeek and Moonshot "relied heavily" on Astra's outputs, citing people with direct knowledge. It does not publish documents, logs, data volumes, or examples of outputs. It does not say how much data was collected, over what period, through which accounts, or what proportion of either model's training it represented. It does not say whether any terms of service were breached, and it does not report any legal action. Readers should not fill those gaps in. A report that a practice happened is not the same as proof of how it happened, or of its legal consequences, and this article does not claim either.

What the report does add is context that predates it. In a blog post on 30 September 2026, OpenAI said it had linked a core part of a distillation campaign that began in July to individuals associated with Moonshot AI. That is a separate, earlier claim by OpenAI about a campaign, not evidence about DeepSeek, and not evidence about Kimi K3.1 Pro specifically. The two should not be merged. The September post shows OpenAI was already investigating distillation linked to Moonshot-associated individuals before the October report; it does not confirm the October report's claims about either company or either model. Keeping the two claims separate is the honest way to read the record as it stands.

The absence of comment is also part of the record, and it should be stated precisely. Reuters reported that OpenAI, DeepSeek and Moonshot AI "did not immediately respond" to its requests for comment. That is not a denial, and it is not an admission. It means that, at publication, none of the three companies had given Reuters a statement. A search for company responses at draft time on 29 October found no new statement from any of the three companies beyond that position, but the position could change, and any response published later should be added to the record rather than assumed.

Credit: Computer chip on a circuit board, via Inkoperscafe

Why distillation has become a flashpoint

Distillation has become a flashpoint because it changes the economics of catching up. Training a frontier model from scratch costs enormous sums in computing, data and engineering time. Training a new model on the outputs of an existing frontier model can reproduce much of its capability more quickly and cheaply, because the hard work of discovering the behaviour has already been done. Firms including Anthropic, OpenAI and Google have argued, in Reuters' summary, that the practice enables rivals to reproduce advanced capabilities more quickly and at lower cost. That argument sits inside a wider US-China technology competition, in which access to advanced chips and frontier models is already restricted, and in which every claim about how a rival model was built carries political weight.

The labs have been building defences, and the report lands in the middle of that effort. Earlier in October 2026, Anthropic said new measures designed to prevent rivals from copying the capabilities of its Claude models had reduced data harvested in what it calls distillation attacks by about 90% to 99%. That figure is Anthropic's own claim about its own systems, not an independent measurement, and it concerns Claude, not Astra. But it shows the direction of travel: the leading labs now treat large-scale output harvesting as a security problem, with detection, rate limits and account controls.

There is a second, quieter point that often gets lost. Distillation is not only a shortcut. It is also how smaller, cheaper models are legitimately made: a large teacher model generates training data, and a smaller student model learns from it, under the provider's own terms and often inside the same company. The line between legitimate distillation and systematic harvesting is drawn by permission and by terms of service, not by the mathematics. Until the method and the permission are known, the fair question is not "did they distil?" in the abstract, but "what did they do, how, and under what terms?" — and the published record does not yet answer it.

Credit: Team working on laptops, via University of Liverpool

What this means for the models named

For the two models named, the report raises a question it cannot answer: how much of their capability came from their own training, and how much from Astra's outputs. Kimi K3.1 Pro and DeepSeek V4.1-Flash 1017 are real products, released on real dates, and their benchmark scores — where independently measured — are what they are, whatever their training data contained. What a training history can change is the legal and commercial standing of the model, the trust of customers, and the willingness of providers to grant access in future. Those consequences depend on facts that have not been published, and on decisions by courts or regulators that have not been taken.

For OpenAI, the report — if accurate — would be the latest sign that its most capable model is also its most copied. For the wider industry, it is another reason to expect tighter controls on API access and more public argument about what model outputs may be used for. None of that requires accepting the report's claims as proven. It follows from the fact that the labs believe the risk is real, and are acting on that belief.

What to watch is straightforward. First, any statement from OpenAI, DeepSeek or Moonshot AI. Second, any evidence published by The Information or another outlet. Third, any legal or regulatory action, of which none has been reported. Until one of those arrives, the correct description of this story is the one in its headline: reported.

Credit: Office workers at desks, via Skills Portal

Key Takeaways

Reported, not proven: Reuters (29 October 2026) reported The Information's account, citing people with direct knowledge, that DeepSeek and Moonshot AI relied heavily on GPT-6 Astra outputs to train their latest models. No evidence has been published and no company has confirmed it.
The models named are Moonshot's Kimi K3.1 Pro (released the week before the report) and DeepSeek's V4.1-Flash 1017 (launched 17 October 2026) — both, according to the report, benefited from distilling Astra.
OpenAI, DeepSeek and Moonshot AI did not immediately respond to Reuters' requests for comment. OpenAI's 30 September post about a distillation campaign linked to Moonshot-associated individuals is a separate, earlier claim — not proof about DeepSeek or K3.1 Pro.
Distillation is a standard technique, used legitimately inside labs; the controversy is about scale, method and permission, not the mathematics. Anthropic says its new measures cut distillation harvesting against Claude by about 90% to 99% (its own claim, unreproduced).
No legal action has been reported, and no amount of data, outputs or breach of terms has been stated. Any company statement or published evidence would change this story and should be added to the record.

A report that two leading Chinese labs trained their newest models on OpenAI's outputs is a serious claim, and it deserves serious treatment: stated as reported, sourced to the outlet that published it, kept separate from earlier claims, and left open to the responses and evidence that have not yet arrived. If the companies respond, or the evidence appears, the story will be updated — and the update will be the part that matters most.

Search This Blog

INORYVA