🔍 Read the full analysis: A Guide To The Distillation Report On China’s AI Models And GPT-6 Astra on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
The Information reported on October 28 that DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra to train recent models. Reuters said the companies did not immediately respond, and the public reporting does not provide evidence about the scale or method. Distillation is a standard training technique; whether access violated OpenAI’s terms or involved deceptive accounts remains unverified.
The Information reported on October 28 that Chinese AI labs DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra to train recently released models. Reuters summarized the report within hours, but the claim has not been publicly confirmed by OpenAI, DeepSeek or Moonshot, and key details about the alleged training and access methods remain unavailable.
The models identified in the reporting are Moonshot’s Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, released October 17. Reuters attributed the central allegation to The Information, which cited people with direct knowledge. The Information’s publicly available article opening described strong demand for Astra, released in early September, and said Chinese AI labs were among a growing customer base seeking to extract its capabilities. The rest of the article is paywalled.
Reuters said the three companies did not immediately respond to requests for comment. The original report, as described in the source material, gives no public figures for how many Astra outputs were used, how they were collected, or what share of either model’s training they represented. Moonshot’s launch post attributed Kimi K3.1 Pro’s gains to substantially more post-training compute and large-scale reinforcement learning; DeepSeek’s update note did not discuss training data. Those company statements do not independently verify or disprove the allegation.
South Korean broadcaster SBS reported additional details on October 29, including that the labs had previously focused on Anthropic’s Claude and shifted toward OpenAI after Anthropic tightened defenses. SBS also quoted an OpenAI spokesperson as saying the company takes the issue seriously and continues working on it. Those points have not been checked against the full, paywalled Information report, so they should be treated as separately attributed reporting rather than independently established facts.
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
The Stakes for Model Training
The dispute matters because distillation can transfer a model’s capabilities into a cheaper or smaller system by training it on another model’s answers. The technique is widely used within AI development: a lab may use its own larger model as a teacher for a more efficient product. The controversy here is not the technique by itself, but the allegation that companies used a rival’s outputs at scale and may have obtained them in ways that breached service terms.
OpenAI’s terms, according to the source material, prohibit using its outputs to build competing models. If the reported conduct is substantiated, the dispute could raise contract and account-access questions, including whether any accounts were fake or resold. The public record supplied so far does not establish that such methods were used, nor does it establish that a law was broken. Legal conclusions would depend on evidence, applicable agreements and jurisdiction.
There is also a practical commercial question. The source lists Astra 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. On the same independent index cited by the source, their scores are 57, 46 and 32, respectively. These figures suggest lower listed prices alongside lower scores on that index; they do not establish that the products perform equally or that distillation explains their results.
For companies choosing AI systems, the allegation is a reminder to ask about training-data provenance and supplier practices, test models on their own workloads, and maintain alternatives where reliability matters. For users, the source reports no immediate change to the apps. It advises caution with cheap, unofficial API resellers, but does not provide evidence that a particular reseller was involved in this case.
Products Worth Considering
DJI Mic 3 Wireless Microphone System (2 Transmitters, 1 Receiver, Charging Case)
LG UltraGear 32GX850A-B 32-Inch 4K UHD Glossy OLED Gaming Monitor
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How Distillation Became a Dispute
Distillation trains a student model on responses from a teacher model. The student learns to produce similar answers, potentially at lower cost. A U.S. government advisory described distillation as a legitimate and useful AI research technique, according to the source material. The practice becomes contentious when the teacher belongs to a competitor and its service rules restrict using outputs to build rival systems.
The allegation also sits alongside earlier accusations, not proof of the present claim. In September, OpenAI linked a “core cluster” in a July campaign to extract hidden reasoning from its models to individuals it associated with Moonshot AI, according to the source. U.S. security agencies named both DeepSeek and Moonshot in an advisory, while China’s commerce ministry called the accusations baseless. Those statements represent opposing positions and do not independently establish that the two labs used Astra outputs to train the models now in question.
The source also describes a separate Anthropic experiment in which a model trained on full reasoning transcripts from Claude Opus 4.8 scored 80.4% on SWE-bench Verified, compared with 71.8% when trained on summarized transcripts. That is Anthropic’s own test, not independent confirmation of the Astra allegation. It illustrates why access to detailed model outputs can matter, while leaving open what information DeepSeek or Moonshot obtained and used.
Evidence Behind the Allegation
The central claim remains unconfirmed publicly. The Information cited people with direct knowledge, but the available account provides no named sources, disclosed records or response from DeepSeek or Moonshot. OpenAI also did not immediately respond to Reuters, according to the report. The companies’ silence at that point is not confirmation.
It is not clear how many Astra outputs were allegedly collected, whether they included hidden reasoning or only final answers, how the outputs were accessed, or how much they influenced either model. The public launch materials described in the source offer other explanations for model development, but do not settle the question. The source further says the underlying Information article is paywalled, limiting independent review of its sourcing and detail.
No legal finding is reported. Whether any conduct violated a contract, access rules or a law would depend on evidence not presented in the public reporting. The claim about a shift from Anthropic to OpenAI, carried by SBS, also remains an additional report that the source author could not verify from the original article.
Responses and Evidence to Watch
The next developments to watch are direct responses from OpenAI, DeepSeek and Moonshot, along with any additional reporting that makes the alleged scale, collection methods or training effects more specific. Public documentation, technical analysis or a formal complaint could clarify parts of the claim, but none is identified in the source material as having been released.
For now, readers and AI buyers should distinguish the reported allegation from established facts. The releases and their dates, the contents of the public launch notes, and the existence of the Information and Reuters reports are documented in the source; the alleged degree of Astra-derived training and any rule-breaking are not publicly demonstrated there. Any assessment may change if the companies or reporters provide further evidence.
Key Questions
Which models does the report name?
The report names Moonshot AI’s Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, released October 17.
Have OpenAI, DeepSeek or Moonshot confirmed the claim?
No. Reuters reported that the three companies did not immediately respond to its requests for comment. The source material does not provide a public confirmation from any of them.
Is distillation itself improper?
Not inherently. Distillation is a common training method, including for creating smaller models from a lab’s own systems. The reported concern is whether a rival’s outputs were used at scale in breach of service terms or obtained through deceptive access. Those details remain unverified.
Does the report prove that the models copied Astra?
No. It reports an allegation based on sources cited by The Information. The publicly available account does not give training-data figures, collection methods or evidence establishing how much Astra outputs contributed to either model.
What should AI buyers do while the claim is unresolved?
Buyers can ask providers about data provenance and compliance practices, test models on their own tasks, and keep a fallback option for important services. The source also advises caution with cheap, unofficial API resellers, without identifying a specific reseller as involved.
Source: ThorstenMeyerAI.com
