Report: DeepSeek and Moonshot trained new models on OpenAI's GPT-6 Astra

DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra to train their newest models, The Information reports, citing people with direct knowledge of the matter. The report names two specific releases: Moonshot’s Kimi K3.1 Pro and DeepSeek’s upgraded V4.1-Flash 1017, both of which it says benefited from distilling Astra.

The claim is unconfirmed. OpenAI, DeepSeek and Moonshot AI did not immediately respond to Reuters’ requests for comment, and the public summaries of the story include no named executive, filing or internal document. It does land on top of a dispute that has been building for months, including an OpenAI investigation that already pointed at Moonshot.

What The Information reports about GPT-6 Astra and the two labs

The original story comes from The Information and was relayed by Reuters. Reuters picked it up on October 28 by its own timeline, with Indian and other outlets publishing their versions on October 29. The core claim is short: the two Chinese labs leaned heavily on GPT-6 Astra outputs when training their latest systems.

The technique at issue is distillation. In the report’s framing, it means using the responses an advanced model generates to train another model, so a developer can reproduce some of its capabilities without the same level of training from scratch. The approach can cut development costs and speed up model development, which is why it has become contentious among AI companies.

Several details are absent from every public version. Shattered, which reviewed the available coverage, notes that none of it adds a number, a cost or a quote from a named person. There is no published estimate of how many Astra outputs were used, how they were collected, or how much they contributed to either model’s performance. Times of AI likewise notes that the extent of the contribution remains unclear.

Kimi K3.1 Pro and DeepSeek V4.1-Flash 1017, the two models named

Both models are recent, and both are open-weight releases, meaning the trained model files are published for anyone to download and run.

Model	Lab	Release	What the sources say
DeepSeek-V4.1-Flash-1017	DeepSeek	October 17, 2026	Focus on sharper task performance and agentic coding; open-weight and low-cost
Kimi K3.1 Pro	Moonshot AI	October 23, 2026	About 2.8 trillion parameters; score of 55 on the Artificial Analysis Intelligence Index, fourth overall

DeepSeek’s update emphasizes agentic coding, where a model plans and carries out multi-step programming work rather than just suggesting code. Moonshot’s model is a substantial upgrade over a predecessor released only a few months earlier. Unicorn API covered that release in Moonshot launches Kimi K3.1 Pro with hacking skills deliberately curbed.

One detail sharpens the story. Crypto Briefing reports that Kimi K3.1 Pro beat GPT-6 Astra on certain agentic benchmarks, including Terminal-Bench 4.0. If the report is accurate, a model said to have learned from Astra’s outputs now outscores Astra on some tests.

The timeline fits the claim without proving it. GPT-6 Astra was introduced in April 2026, months before either Chinese release, so a teacher model was available well before the students shipped. As Shattered puts it, a timeline that fits is not proof.

What OpenAI and Anthropic have already said about distillation

This is not the first time Moonshot has been linked to distillation of OpenAI models. In a September 30 blog post, OpenAI said it had linked a core part of a distillation campaign that began in July to individuals associated with Moonshot AI.

Times of AI summarizes the details OpenAI gave about that campaign:

• The activity began on July 1, with spikes on July 24 and 25.
• OpenAI recorded 16,000 requests using a relevant extraction pattern from more than 4,000 users.
• Related activity spanned a cluster of more than 15,000 users, which OpenAI said it had fully disrupted by July 28.
• The figures describe attempted extractions; OpenAI said they do not mean every attempt succeeded.
• The campaign sought to expose protected reasoning through manipulated model interactions, without breaking encryption, compromising a database or accessing stored user conversations.

OpenAI said it responded by restricting accounts, strengthening technical controls and improving monitoring for coordinated activity.

Anthropic has its own history with both labs. In February 2026, it accused Moonshot and DeepSeek of running distillation attacks on its Claude models and alleged the labs used fraudulent accounts to extract outputs. In an October 12 disclosure, Anthropic said its guardrails detect suspicious requests, restrict access to protected reasoning and make repeated extraction attempts harder, and that it had banned tens of thousands of accounts tied to distillation.

The reported effect of those safeguards differs slightly between outlets. ETEnterpriseAI and TechGig report that Anthropic’s new measures cut data harvested in distillation attacks by about 90 to 99 percent, while Times of AI gives the range as 92 to 99 percent.

The US government’s position on distillation by Chinese labs

The dispute has a government dimension that predates this report. In April 2026, the US State Department circulated a diplomatic cable to missions worldwide that named DeepSeek, Moonshot AI and MiniMax. According to officials, the cable was meant to:

warn of the risks of utilising AI models distilled from U.S. proprietary AI models, and lay the groundwork for potential follow-up and outreach by the U.S. government.

Diplomatic staff were told to raise “concerns over adversaries’ extraction and distillation of U.S. A.I. models” with foreign governments and partners. The Chinese Embassy in Washington has previously rejected claims that Chinese companies ignore intellectual property rules, saying Beijing “attaches great importance to the protection of intellectual property rights”.

Times of AI also notes that Chinese developers have faced resource restrictions, including limits on access to advanced chips. That is part of why the cost advantages of distillation attract attention, though the same outlet stresses that distillation is not the only factor behind Chinese models’ progress; gains in training and inference efficiency also contribute.

Why distillation is hard to prove, and where the line sits

Distillation itself is routine. Labs use it on their own models to shrink a large system into a cheaper, faster one, and OpenAI and Anthropic have both described using it in their own development. The contested version is distilling a rival’s proprietary model without authorization.

Mechanically, the two look the same. A developer sends a large set of prompts to the teacher model, records the responses, and uses those prompt-and-response pairs to train the student. The student never sees the teacher’s weights or code; what moves is text. That makes the practice hard to prove from the outside. Industry specialists quoted in April coverage of the State Department cable made the same point: determining exactly how a model was trained is technically complex.

Whether any particular use breaks a provider’s rules depends on that provider’s terms of service, and nothing in the public reporting says which side of the line the reported use of Astra falls on. People familiar with such efforts told the publication that distillers may use multiple accounts to gather enough outputs while avoiding detection, which forces providers to separate normal customer activity from systematic extraction.

What the report means for developers using these models

For teams building on Kimi K3.1 Pro or DeepSeek V4.1-Flash 1017, nothing about the models changes today. The weights are the same as before the report. What changes is the risk picture around them.

• Provenance questions. The State Department cable explicitly warns partners about the risks of using models distilled from US proprietary models. Teams selling to governments or regulated customers may face more questions about where their underlying models came from.
• Durability of the cost advantage. Crypto Briefing argues that if a lab’s edge depends partly on a competitor’s outputs, investors have to ask how durable that edge is. The same applies to builders choosing a model family for the long term: tighter anti-extraction controls at US labs could, in principle, affect how future versions are trained.
• Collateral effects of anti-distillation controls. Anthropic acknowledged that its measures can occasionally affect legitimate users. Developers who send large, repetitive volumes of requests to frontier APIs for their own evaluation or data work should expect closer monitoring of coordinated activity, the kind OpenAI says it has strengthened.

There are also financial stakes. Moonshot AI’s valuation reportedly reached around $50 billion ahead of a planned Hong Kong IPO, and DeepSeek has reportedly been valued higher in recent funding rounds. Distillation allegations are the kind of issue prospective shareholders tend to scrutinize.

What to watch next

• Company responses. OpenAI, DeepSeek and Moonshot AI had not responded at the time of the report. A denial, a confirmation, or a new OpenAI disclosure naming DeepSeek would change the evidence picture.
• Underlying evidence. So far the public record is a sourced report with no volumes, costs or documents. Any technical detail, such as request counts like those OpenAI published for the July campaign, would make the claim testable.
• Government follow-up. The April cable described laying “the groundwork for potential follow-up and outreach by the U.S. government”. This report gives Washington two new, specific models to point at.
• Moonshot’s IPO. A Hong Kong listing at around $50 billion would invite detailed questions about training practices.
• More provider safeguards. Times of AI expects distillation to remain a key issue as companies strengthen guardrails against large-scale extraction.