🔍 Read the full analysis: A Question Of Distillation: China’s New AI Models And GPT-6 Astra on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- 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 OpenAI’s GPT-6 Astra outputs to train recently released models. The claim is based on unnamed sources; the three companies had not confirmed or denied it in Reuters’ account, and the public record does not explain the scale or method.
The Information reported 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, a claim that could raise questions about model-training practices and the terms governing access to commercial AI systems. The allegation is based on unnamed sources, and the companies had not publicly confirmed or denied it in Reuters’ account.
Reuters summarized The Information’s report as saying that the models benefiting from Astra outputs were Moonshot’s Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, released October 17. The Information cited people with direct knowledge, according to the report. Its article’s public opening described Chinese AI labs as a growing customer base for Astra; the remainder was behind a paywall, limiting what could be independently assessed from the public material.
The available reporting does not state how many Astra outputs were used, what prompts produced them, how the labs obtained access, or how much those outputs contributed to either model. Reuters reported that OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment. The source material also says the labs’ own release notes offer different explanations or no explanation: Moonshot credited Kimi’s gains to increased post-training compute and reinforcement learning, while DeepSeek’s update note did not detail its training data.
A separate account by South Korea’s SBS, published October 29, attributed further details to the same report, including that Chinese labs had previously focused on Anthropic’s Claude and shifted toward OpenAI after Anthropic tightened defenses. SBS also reported a statement from an OpenAI spokesperson saying the company takes the issue seriously and continues working on it. Those details could not be checked against The Information’s full paywalled report in the source material.
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 Access Matters
If the report is accurate, it would put a spotlight on the boundary between ordinary distillation and using a rival’s model outputs at scale in ways that breach its terms. Distillation—training a smaller model to reproduce answers from a larger one—is a common technique in AI development. The U.S. government advisory cited in the source material calls it a legitimate and useful research method. The disputed questions are who owns the teacher model, the scale of copying, and whether access complied with the provider’s rules.
The commercial stakes are visible in the prices and benchmark figures provided by the source. GPT-6 Astra is listed at $10 per million input tokens and $50 per million output tokens; Kimi K3.1 Pro at $2 and $10; and DeepSeek’s model at $0.30 and $1.20 during peak hours. On the same independent index, their scores are reported as 57, 46 and 32, respectively. These figures indicate lower listed prices alongside lower scores on that index; they do not establish that one model is better for every task or explain why the models perform differently.
For businesses choosing models, the report is a reason to ask vendors about training-data provenance and keep alternatives available, not proof that either named model violated a rule. For ordinary users, the source says nothing changes in the apps today. It advises caution with inexpensive, unofficial API resellers, but does not provide evidence that a particular reseller was involved.
Products Worth Considering
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
Logitech MX Brio Ultra HD 4K Webcam, Graphite
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Distillation and Earlier Allegations
In distillation, a larger model acts as a “teacher”: it produces answers to prompts, and a smaller “student” model is trained to imitate them. Labs commonly use the method on their own systems to make models less expensive to run. The technique itself is not evidence of misconduct; the question in this report is whether outputs from a competitor’s model were used and, if so, whether the way they were obtained or used broke contractual terms.
The report follows earlier public allegations described in the source material. In September, OpenAI connected a “core cluster” of a July campaign to extract hidden reasoning from its models to individuals associated with Moonshot AI, and U.S. security agencies named both DeepSeek and Moonshot in an advisory. China’s commerce ministry called the accusations baseless. The supplied account also refers to earlier allegations involving activity on Anthropic’s Claude, but does not provide enough detail to independently establish the scope or responsibility for that activity.
OpenAI’s terms, as summarized by the source, prohibit using its outputs to build competing models. That makes the current allegation primarily a contract and access question, with possible concerns about fraud if access was obtained through fake or resold accounts. The report does not establish that such accounts were used, or that any law was broken.
What the Public Record Cannot Show
The allegation remains unverified in the public record. The core account relies on unnamed sources, and key evidence—including logs, datasets, account records and a detailed explanation of the alleged training process—has not been made public in the supplied material. It is also unclear whether the reported use, if it occurred, involved direct access by the labs, third-party resellers, or another route.
No company has publicly confirmed the allegation in the Reuters account cited here. The available information does not establish whether Astra’s outputs materially changed the released models, whether any contract was breached, or whether the labs’ own descriptions of their training processes conflict with the report. The legal characterization is also unsettled: a possible breach of a service agreement is not, by itself, proof of a criminal offense.
Responses and Evidence to Watch
The next useful developments would be direct responses from OpenAI, DeepSeek and Moonshot, along with any more detailed reporting that identifies the methods, scale and source of the alleged outputs. Public documentation or technical evidence could help distinguish ordinary model training from systematic use of a rival’s service, but no such evidence is described in the material provided.
Companies evaluating these models can ask providers about training provenance, account controls and contract compliance, then test systems against their own requirements. Those steps are prudent procurement checks, not a finding that the report’s allegation is true. Users should rely on official services and verified API channels while the claims remain unresolved.
Key Questions
Which models are named in the report?
The Information’s report, as summarized by Reuters, names Moonshot AI’s Kimi K3.1 Pro and DeepSeek-V4.1-Flash-1017. Their release dates were reported as October 23 and October 17, respectively.
Has either AI lab confirmed using GPT-6 Astra outputs?
No confirmation is cited in the available reporting. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment.
Is distillation itself improper?
No. Distillation is a standard way to train smaller models using a larger model’s outputs. The concern raised by the report is whether a rival’s outputs were used at scale and whether access or use violated the provider’s terms.
What evidence is public so far?
The central claim comes from unnamed sources cited by The Information and summarized by Reuters. The public material does not disclose the volume of outputs, access method or training contribution, and the named companies have not confirmed the allegation in the cited account.
Does the report change what users should do?
The source material says there is no immediate change to the apps. Businesses may ask vendors about training-data provenance and maintain alternatives; users should be cautious with unofficial API resellers while the claim remains unresolved.
Source: ThorstenMeyerAI.com
