🔍 Read the full analysis: What A Distillation Report Can—and Can’t—Tell Us About China’s AI Models on ThorstenMeyerAI.com
Get smart everyday buys delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
The Information reported on Oct. 28 that DeepSeek and Moonshot AI relied heavily on OpenAI’s GPT-6 Astra outputs to train recently released models. The report cites people with direct knowledge, but its detailed evidence is paywalled, and the companies had not confirmed or denied the account in Reuters’ report. Distillation is a common training method; whether it involved unauthorized access or breached terms remains unclear.
The Information’s report, by Jing Yang, Juro Osawa and Qianer Liu, cited people with direct knowledge of the matter. Reuters summarized the claim as saying DeepSeek and Moonshot’s latest models “relied heavily” on Astra outputs. The named releases are Moonshot’s Kimi K3.1 Pro, released Oct. 23, and DeepSeek’s V4.1 Flash update, launched Oct. 17.
The original article’s public opening described Chinese AI labs as a growing customer base for Astra, which The Information called OpenAI’s most powerful released model. The rest of its reporting is behind a paywall. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its inquiries. That means the allegation has not been publicly confirmed by the companies involved, and the available reporting does not establish the volume of outputs used, the prompts sent or the accounts involved.
A separate report by South Korean broadcaster SBS on Oct. 29 said the sources described a shift from using Anthropic’s Claude to OpenAI models after Anthropic tightened defenses. SBS also quoted an OpenAI spokesperson saying the company takes the issue very seriously and continues working on it. The original reporting could not be independently checked against the paywalled article for this account, so those details should be treated as SBS’s summary, not as independently verified findings.
The companies’ own public descriptions do not resolve the question. Moonshot’s launch post credited Kimi K3.1 Pro’s gains to “substantially larger post-training compute” and large-scale reinforcement learning. DeepSeek’s update note did not disclose training data. Neither statement, on its own, confirms or rules out use of Astra outputs.
Infographic 01 · Distillation in plain words
How distillation works, and where it crosses a line.
A big model teaches a small one by example. The technique is ordinary; the dispute is about whose model is the teacher and how the answers were obtained.
Send the teacher prompts
A strong “teacher” model gets a huge number of carefully chosen tasks.
Keep answers and reasoning
The step-by-step working is worth far more than the final answer alone.
Teach a student to imitate
A smaller or cheaper “student” model learns to reproduce the teacher’s moves.
Normal practice
The U.S. advisory calls it “a legitimate and useful technique”.
- A lab distills its own large model
- Goal: smaller, faster, cheaper versions
- Open models used within their licence
Where the fight starts
What OpenAI, Anthropic and U.S. agencies allege.
- The teacher is a rival’s model
- Industrial scale: millions of exchanges
- Fake, resold or stolen accounts and keys
- Breaks terms that ban building competing models
Why hidden reasoning is the prize
Anthropic’s own test: a model trained on Claude Opus 4.8 transcripts, scored on SWE-bench Verified (coding).
That is why providers now hide or summarize reasoning, and why extraction attempts target it.
Distillation is a student copying a top tutor’s worked solutions. Legal in a classroom. Contested when the tutor never agreed.Thorsten Meyer · my reading
Sources: NSA/CISA/FBI advisory AA26-251A, 8 Sep 2026; Anthropic, 12 Oct 2026 (vendor’s own experiment). Analogy and framing: my reading.
Powered by Thorsten Meyer AI
Why Model Training Provenance Matters
The report matters because training data and model outputs can be commercially valuable. If a lab uses a rival’s paid system at scale to reproduce capabilities, the dispute may involve contract terms, access controls and competition. But whether those conditions occurred here is not established by the public information available.
The source material lists Astra pricing at $10 per million input tokens and $50 per million output tokens, compared with $2 and $10 for Kimi K3.1 Pro, and $0.30 and $1.20 for DeepSeek’s model at peak hours. It also gives scores of 57, 46 and 32, respectively, on the same independent index. These are attributed figures, not proof of how the models were trained; the index and its testing details are not described in the supplied reporting. The comparison suggests different price and score positions, not equivalent performance.
For organizations choosing models, the practical issue is broader than this allegation: training provenance can affect confidence in a provider’s methods and the durability of access. Buyers can ask vendors about data sources and training processes, test models on their own workloads and maintain a fallback provider. The report does not show that users need to change the apps they use today.
Products Worth Considering
HP Sprocket Photo Booth Machine Instant Color Photo Printer with 10.1 Touchscreen, LED Ring Light, 250+ Frames, 3×4 ZINK Paper, QR Code Shar
Dell 34-Inch S3425DW Curved WQHD USB-C Monitor
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
As an affiliate, we earn on qualifying purchases.
Distillation and Earlier Allegations
Distillation is a standard way to train a smaller model using responses from a larger “teacher” model. A student model is exposed to many answers and trained to reproduce patterns in them. The technique is used by labs to make capable systems that may cost less to run. The U.S. government advisory described in the source material calls distillation “a legitimate and useful technique in AI research.” The controversy is not the method alone, but whether it is applied to a competitor’s model at scale and through access that violates rules.
The source material also describes earlier allegations and public disputes. In February 2026, Anthropic said DeepSeek, Moonshot and MiniMax had generated more than 16 million exchanges with Claude through accounts it associated with those labs; the provided text cuts off before giving the full account or its outcome. In September, OpenAI tied what it called a “core cluster” of a July campaign to extract hidden reasoning to individuals associated with Moonshot, according to the source. U.S. security agencies also named both labs in an advisory, while China’s commerce ministry called the accusations baseless. These are previous claims and responses, distinct from proof of the October report’s specific allegation.
OpenAI’s terms, as described in the source material, prohibit using its outputs to build competing models. Whether that restriction was breached depends on facts not yet public, including the applicable terms, how access was obtained and what outputs were used. Any legal conclusion would require more evidence and the relevant circumstances.
“DeepSeek and Moonshot “relied heavily on outputs” from GPT-6 Astra to train their latest models.”
— The Information, as summarized by Reuters
Evidence and Access Remain Unclear
The available public reporting does not specify how many Astra outputs were allegedly used, which capabilities they covered, or whether the models were trained on ordinary answers, hidden reasoning or both. It also does not identify the accounts or access methods at issue. Without those details, readers cannot independently assess the scale or verify the central claim.
There is also no public confirmation or denial from DeepSeek or Moonshot in the Reuters account, and no detailed response from OpenAI there. The distinction between legitimate experimentation and a terms breach may turn on contract language and the means of access; possible fraud has been raised as a general concern, not established in this case. The report alone does not show that either lab broke a law.
Responses and Evidence to Watch
The next developments to watch are direct responses from OpenAI, DeepSeek and Moonshot, along with any further reporting that makes the evidence and methods more specific. Public details about access controls, account use, training sources or a formal investigation could clarify whether the allegation concerns ordinary model use, prohibited model-building or another issue.
For now, the confirmed development is that The Information published the allegation and Reuters reported it, not that the companies have established or resolved the underlying facts. Buyers and users can follow provider disclosures and use official services rather than low-cost, unofficial API resellers, whose access and reliability may be difficult to verify. No immediate change to the named consumer products was reported.
Key Questions
What did the report allege?
The Information reported that DeepSeek and Moonshot AI relied heavily on GPT-6 Astra outputs to train recent models, including Kimi K3.1 Pro and DeepSeek V4.1 Flash. The companies had not confirmed the claim in Reuters’ account.
What is distillation?
Distillation trains a smaller model using responses from a larger one as examples. It is a common AI training technique; the dispute here concerns alleged use of a rival’s outputs and whether access or use violated applicable terms.
Has unauthorized access been proven?
No. The public reporting provided does not describe the accounts, access methods or underlying evidence in enough detail to establish that access was unauthorized or that a law was broken.
Do users need to change AI apps now?
The source material reports no immediate change to the apps. Users can stick to official services and avoid unofficial resellers if they cannot verify where access comes from.
Source: ThorstenMeyerAI.com
