AI Agent UpdateReport: DeepSeek and Moonshot trained their latest models on OpenAI's GPT-6 Astra outputs
Reuters, citing The Information, reports that two of China's leading AI labs relied heavily on outputs from OpenAI's newest model to train their own — a distillation allegation both labs have not yet answered.
The News
Reuters reported on October 28, citing The Information and people with direct knowledge, that Chinese developers DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's GPT-6 Astra model to train their latest models. The models are Moonshot's Kimi K3.1 Pro, released last week, and DeepSeek's V4.1-Flash 1017, launched on October 17. This is a report, not a confirmed finding: it rests on The Information's sourcing, and OpenAI, DeepSeek and Moonshot did not immediately respond to Reuters' requests for comment. Nothing in this story should be read as an admission by either lab.
Distillation, in plain language, is learning from a teacher's answers instead of doing the homework yourself. Training a frontier AI model from scratch takes enormous computing power and data. A shortcut is to ask an existing powerful model thousands or millions of questions, collect its answers, and train a new model on those answers — transferring much of the teacher's ability at a fraction of the cost. Model makers generally forbid this in their terms of service, because it lets a rival copy the capability without paying the training cost. That is why the allegation is a flashpoint in US-China AI competition, not just a technical footnote.
The context in the same report strengthens the allegation's plausibility without confirming it. OpenAI said in a September 30 blog post that it had linked a core part of a distillation campaign begun in July to individuals associated with Moonshot AI. Separately, Anthropic said earlier in October that its anti-distillation measures had cut harvested data by about 90-99%. Neither fact proves the specific claim about GPT-6 Astra outputs, but together they show that distillation is an active, policed problem for frontier labs — and that OpenAI had already pointed at Moonshot's orbit once before.
A model trained on a rival's answers is not a rival — it is a copy with good marketing.
Why It Matters
If the report is accurate, it changes how the AI race should be scored. DeepSeek and Moonshot have been presented as evidence that China can match frontier AI capability at much lower cost. If part of that capability was distilled from OpenAI's model, then the cost comparison is misleading: the teacher paid the training bill, and the student copied the answers. That does not make the students' engineering trivial — distillation still requires real skill — but it does mean their models are not independent proof that frontier capability can be built cheaply from scratch.
For OpenAI and its partners, the stakes are commercial as well as geopolitical. OpenAI's models are its product; if rivals can harvest their outputs and train competitors on them, OpenAI is effectively subsidising its own competition. Microsoft is the read-through ticker here as OpenAI's major partner and investor, and Nvidia as the compute supplier whose chips power both sides of the race. Both are exposed to a world where model capability spreads faster than model revenue, because the capability can be copied even when the training cost cannot be avoided by the originator.
The policy consequences are already visible. Anthropic's claim that its anti-distillation measures cut harvested data by 90-99% shows frontier labs are now building technical defences, not just legal ones, against output harvesting. Expect that trend to accelerate: rate limits, output watermarking, and legal terms aimed at distillation are becoming part of the competitive moat, alongside raw model quality. The labs that protect their outputs best may keep their lead longest, regardless of who trains the cleverest model.
Analyst's Take
Treat this as a serious, well-sourced allegation — not a verdict. Reuters reporting The Information, plus OpenAI's own September blog post linking a distillation campaign to Moonshot-associated individuals, makes this more than a rumour. But neither lab has responded, and no primary evidence has been published. The Bearish verdict reflects what the allegation means if true: a crack in the story that Chinese labs are matching frontier AI on their own, and a warning that model capability is easier to copy than to create.
What to watch is the response. If DeepSeek or Moonshot deny the claim with specifics, or if OpenAI publishes evidence, the story moves from allegation to fact in one direction or the other. If the labs stay silent and frontier labs tighten their anti-distillation defences further, the market will read silence as its own answer. Either way, the era of treating every new model release as an independent achievement is over — the first question for the next release will be what it was trained on.
By The Numbers
FAQ
What is distillation?
Training a new AI model on the outputs of an existing model, rather than on original data. You ask the powerful model many questions, collect its answers, and teach the new model to imitate them. It transfers much of the teacher's ability at a fraction of the training cost, and most model makers forbid it in their terms.
Is this proven?
No. It is a report by Reuters citing The Information and people with direct knowledge. OpenAI, DeepSeek and Moonshot did not immediately respond to Reuters' requests for comment. OpenAI's September 30 blog post about a Moonshot-linked distillation campaign is supporting context, not proof of this specific claim.
Is this a repeat of the Kimi K3.1 Pro story?
No. The Kimi K3.1 Pro launch itself was published here on October 24. The new event is the October 28 allegation about how that model — and DeepSeek's V4.1-Flash 1017 — were trained.
Why are Microsoft and Nvidia the tickers?
DeepSeek and Moonshot are private, so there is no stock to point at. Microsoft is OpenAI's major partner and investor, and Nvidia supplies the computing hardware for both sides of the AI race. They are read-through tickers, not companies accused of anything.
What happens if the allegation is confirmed?
Expect legal, technical and policy consequences: tougher terms of service, stronger anti-distillation defences like rate limits and watermarking, and a reassessment of how independently Chinese labs can match frontier models. It would also sharpen US-China technology competition, where model capability is already a strategic issue.
👁
Source: Reuters via WNCY