29 October 2026 — Chinese artificial intelligence developers DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's GPT-6 Astra model to train their latest models, The Information reported on Wednesday, 28 October 2026, according to Reuters, in a claim that has reignited the industry's bitter dispute over a technique known as distillation.
Illustration: Two stylised artificial intelligence language models exchanging streams of text and data, with a larger model passing glowing tokens to a smaller model being trained.
What Happened
According to the report, cited by Reuters on 29 October 2026, both companies benefited from distilling GPT-6 Astra, OpenAI's powerful flagship model, into their newest releases. Moonshot AI's Kimi K3.1 Pro, which was released last week, and DeepSeek's upgraded V4.1-Flash 1017 model, which was launched on 17 October 2026, are the two models named in the report.
Reuters said the report cited people with direct knowledge of the matter. It is important to be clear about the status of the claim: this is a report, not a finding published by any of the companies involved. OpenAI, DeepSeek and Moonshot AI did not immediately respond to Reuters' requests for comment, Reuters said, so none of the three companies has confirmed or denied the specific account in the reporting so far.
Distillation, in this context, means training a newer model on the outputs of an established, more capable model. Instead of starting only from raw text, images or other training data collected independently, a developer feeds a frontier model a large number of prompts, collects its answers, and uses those answers as training examples for a smaller or newer system. The technique can transfer reasoning patterns, style and problem-solving approaches from the teacher model to the student model.
Leading American AI laboratories argue that this crosses a line when it is done without permission. Their case, as described by Reuters, is that distillation allows rivals to reproduce advanced capabilities more quickly and at lower cost than building them from scratch, effectively free-riding on the enormous investment in data, computing power and research that went into the original model. Critics of that view, in the wider industry debate, argue that learning from outputs is a long-established way for models to improve, and that the boundaries between legitimate training, imitation and misuse are still being worked out in company terms of service, licensing and, increasingly, in courts and regulation.
Background
The new report does not arrive out of nowhere. It follows a series of public warnings from the major laboratories about extraction and copying of their models, and it lands in the middle of an intensifying technological competition between the United States and China.
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, according to Reuters. That earlier disclosure did not, in the Reuters account, name the same models or make the same specific claim as the new report, but it put Moonshot AI at the centre of OpenAI's concerns about large-scale extraction of its outputs.
Anthropic, another leading American laboratory, has also moved publicly on the issue. Earlier in October 2026, Anthropic said new measures designed to prevent rivals from copying the capabilities of its Claude AI models had reduced the data harvested in so-called distillation attacks by about 90 percent to 99 percent, Reuters reported. The figure is Anthropic's own assessment of its defences, not an independent audit, but it shows how seriously the laboratories now treat output extraction as a security problem, not merely a commercial one.
Illustration: Two abstract robot figures facing each other across a digital map with data centre buildings and server racks, symbolising US-China technology competition.
The timing matters for another reason. Both models named in the report were released in October 2026, and both have been presented as evidence that Chinese developers are closing the gap with American frontier laboratories. Kimi K3.1 Pro is Moonshot AI's latest flagship, while DeepSeek's V4.1-Flash 1017 is an upgraded, faster variant in a model family that first drew global attention for achieving strong results at a fraction of the training cost claimed by its American rivals. If the new report is accurate, at least part of that rapid progress would have been built, directly or indirectly, on OpenAI's work.
That question — how much of a model's capability is genuinely original, and how much is borrowed — has become central to how the AI race is judged. Benchmark scores and release notes tell the public what a model can do, but they rarely reveal in detail what it was trained on. Companies guard their training data closely for commercial and legal reasons, which means outside observers often have to rely on reporting, leaks and the companies' own disclosures to piece together the story.
The broader policy backdrop adds further pressure. Washington has spent recent years restricting exports of advanced chips and chipmaking equipment to China, on the grounds that they could support military and surveillance applications. Beijing, for its part, has promoted its domestic AI industry and, in October 2026, proposed new export controls of its own covering AI technologies. In that environment, a dispute about model outputs is not only a commercial quarrel between companies: it feeds into questions about whether controls on hardware can be effective when software capabilities can be transferred through an application programming interface, one prompt at a time.
Why It Matters
First, the report, if confirmed, would sharpen the debate about the rules of the AI industry. At present, there is no settled global standard for when training on another model's outputs is acceptable. Some uses are openly commercial and licensed: a company may pay to use a frontier model and, within the terms of that licence, fine-tune a smaller model for a specific task. Other uses are explicitly prohibited by the terms of service of major providers, which typically bar using their outputs to train a competing model. The dispute is therefore partly about contract and consent, and partly about what the industry as a whole should tolerate.
Second, it affects how investors, customers and governments judge Chinese AI progress. DeepSeek and Moonshot AI have both been held up as proof that capable AI can be built without the largest budgets in the world. That narrative matters for fundraising, for hiring, for export-control policy and for national prestige. A finding that their latest models leaned heavily on a rival's outputs would not erase their engineering work — distillation still requires substantial expertise to do well — but it would change the story about how independent that progress has been.
Third, it is likely to accelerate the defensive measures already under way at the major laboratories. Anthropic's reported 90 to 99 percent reduction in harvested data shows what is possible with rate limits, detection of automated extraction patterns and other technical controls, but such defences also risk blocking legitimate users, researchers and developers who rely on high-volume access. The laboratories must balance openness, which has driven much of AI's recent progress, against protection of the assets that make their businesses valuable.
Finally, the episode illustrates how quickly the AI race is becoming a contest over rules as well as technology. Chips, data centres and model weights matter, but so do terms of service, licensing, export controls and the norms that govern what one laboratory may learn from another. The outcome of those rule-setting battles may shape the industry as much as the next model release.
What Happens Next
The immediate question is whether any of the three companies will respond in detail. OpenAI, DeepSeek and Moonshot AI have so far not commented to Reuters on the specific report, and a fuller response — a denial, a confirmation, a legal warning or a technical rebuttal — would be the first test of how seriously the claim is taken by the parties themselves.
Watch also for any change in how the major laboratories police their services. If distillation is becoming harder to detect or easier to do at scale, further technical restrictions, new licensing terms or legal action are plausible next steps, following the pattern already set by Anthropic's October measures and OpenAI's September disclosure.
Illustration: A digital shield protecting a large language model server, with intercepted data extraction attempts shown as blocked streams.
Regulators and policymakers may also take an interest, particularly in Washington and Beijing, where AI competition is already framed as a matter of economic and national security. Any official response, however, would need to grapple with a difficult fact: distillation is hard to prove from the outside, and even harder to regulate without also restricting the ordinary, legitimate use of AI models by millions of developers and researchers.
For users, the practical impact is likely to be indirect. Models may become more expensive to access at high volume, terms of use may tighten, and the provenance of a model's capabilities — where it learned what it knows — may become a more prominent part of how products are marketed and judged. What will not change overnight is the capability of the models already released: Kimi K3.1 Pro and DeepSeek's V4.1-Flash 1017 are already in the hands of users, whatever their training history turns out to have been.
Sponsored: explore today's featured sponsored offers
Frequently Asked Questions
What did The Information report?
According to Reuters' account of the report, DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's GPT-6 Astra model to train their latest models, Moonshot's Kimi K3.1 Pro and DeepSeek's V4.1-Flash 1017.
What is distillation?
Distillation is a training technique in which a newer model learns from the outputs of an established, more capable model, rather than only from independently collected training data. It can transfer reasoning patterns and problem-solving approaches from the teacher model to the student model.
Have the companies confirmed the report?
No. Reuters said OpenAI, DeepSeek and Moonshot AI did not immediately respond to its requests for comment, so the claim remains a report based on people with direct knowledge, not a confirmed admission by any company.
Why do AI laboratories object to distillation?
OpenAI, Anthropic and Google have argued, according to Reuters, that the practice allows rivals to reproduce advanced capabilities more quickly and at lower cost, undermining the investment that went into building the original model.
What have the laboratories already done about it?
OpenAI said on 30 September 2026 that it had linked part of a distillation campaign that began in July to individuals associated with Moonshot AI, and Anthropic said earlier in October that new countermeasures had cut data harvested in distillation attacks against its Claude models by about 90 percent to 99 percent.
Sources
- Reuters (Disha Mishra), "DeepSeek, Moonshot used OpenAI's GPT-6 Astra to train latest models, The Information reports," 29 October 2026.
- The Information, report on DeepSeek and Moonshot AI distillation of GPT-6 Astra, 28 October 2026, as reported by Reuters.
- OpenAI blog post, 30 September 2026, on a distillation campaign linked to individuals associated with Moonshot AI, as reported by Reuters.
- Anthropic, October 2026 measures against distillation of Claude models, as reported by Reuters.
Comments
Post a Comment