🔍 Read the full analysis: The Distillation Report On China’s AI Models, Explained on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
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 claim is based on unnamed sources, and the public record does not include evidence about the scale or method; the companies had not responded to Reuters when it published its account.
The Information reported Oct. 28 that Chinese AI labs DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra to train models released this month, a claim that could raise questions about access to a rival’s technology and compliance with OpenAI’s terms. The report cites people with direct knowledge, but the public evidence is limited and the companies had not confirmed or denied the allegation in Reuters’ account.
Reuters summarized the report as saying that Moonshot’s Kimi K3.1 Pro and DeepSeek’s V4.1 Flash update benefited from distilling Astra. The models were released on different dates: Kimi K3.1 Pro on Oct. 23 and DeepSeek-V4.1-Flash-1017 on Oct. 17, according to the source material. The Information’s full article is paywalled, and its public opening describes strong demand for Astra since its early-September release, including interest from Chinese AI labs.
Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to requests for comment. A separate account by South Korean broadcaster SBS attributes to an OpenAI spokesperson a statement that the company takes the issue seriously and continues working on it. The source material says that detail could not be checked against The Information’s original report, so it remains a separately reported statement rather than independent confirmation of the allegation.
The available accounts do not provide usage totals, training methods, account details or evidence showing how much either model’s performance came from Astra outputs. Moonshot’s launch post attributed Kimi K3.1 Pro’s gains to increased post-training compute and large-scale reinforcement learning. DeepSeek’s update note, according to the source material, did not describe its training data. Those statements do not settle whether Astra outputs were used.
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 Access
If the report is accurate, the dispute is about more than whether AI models can learn from one another. It could concern how a provider’s outputs were obtained, whether they were used at industrial scale, and whether that use breached the provider’s terms. OpenAI’s terms prohibit using its outputs to build competing models, according to the source material. Whether any particular conduct violated those terms would depend on facts not available in the public accounts and on the applicable legal analysis.
The report also matters to customers evaluating AI products. The source material 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 there, the models scored 57, 46 and 32, respectively. The figures describe different prices and scores, not proof of how the models were trained or a like-for-like measure of value.
For businesses, the immediate practical issue is model provenance and supplier risk: buyers may want clearer information about training sources and a fallback provider. The report does not establish that users’ current applications are affected. The source material advises users to avoid cheap, unofficial API resellers, but offers no evidence that a particular reseller was involved in this case.
Products Worth Considering
Bose Lifestyle Ultra Wireless Home Speaker
Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
As an affiliate, we earn on qualifying purchases.
Distillation and Earlier Allegations
Distillation is a standard training technique: a smaller student model learns from outputs generated by a larger teacher model. Labs can use it to produce less costly models based on their own systems. The U.S. government advisory described distillation as a “legitimate and useful technique in AI research,” according to the source material. The controversy arises when a company is alleged to use a competitor’s outputs in ways restricted by the provider’s terms or obtain them through deceptive access.
The Astra report follows earlier allegations and warnings. In September, OpenAI said a “core cluster” from a July campaign aimed at extracting hidden reasoning from its models was linked to people associated with Moonshot AI, according to the source material. U.S. security agencies also named DeepSeek and Moonshot in an advisory. China’s commerce ministry called the accusations baseless. Those prior claims are relevant background, but they do not independently prove that either lab used Astra outputs to train the models named in the new report.
The source material also describes a February 2026 Anthropic account involving large volumes of exchanges with Claude, and says OpenAI had previously raised concerns about attempts to extract model reasoning. These accounts point to broader concern among AI providers about competitors or other users reproducing model capabilities. They are not a substitute for evidence about the specific training of Kimi K3.1 Pro or DeepSeek-V4.1-Flash-1017.
“DeepSeek and Moonshot “relied heavily on outputs from OpenAI’s powerful GPT-6 Astra model to train their latest models.””
— The Information, as summarized by Reuters
Evidence Behind the Training Claim
The allegation remains unconfirmed by the three companies in the Reuters account. The Information’s report relies on unnamed sources, while the public material available here does not disclose the amount of Astra output used, the prompts or methods involved, or how any access was obtained. It is also unclear whether the alleged use, if it occurred, involved direct access by either lab or intermediaries such as resellers.
There is no public breakdown showing what proportion of either model’s training or capabilities came from Astra, and the models’ launch statements do not resolve that question. The source material does not include a response from DeepSeek or Moonshot to the specific allegation. No conclusion about a legal violation can be drawn from the reported information alone.
Responses and Further Evidence
The next developments to watch are responses from OpenAI, DeepSeek and Moonshot, along with any additional reporting that identifies the evidence, scale or access methods behind the allegation. The public account also leaves open whether OpenAI may pursue account restrictions, contractual remedies or other action; no such response is confirmed in the source material.
For now, the report should be treated as an allegation rather than a verified account of the models’ training. Companies choosing AI providers can ask vendors about training-data provenance and test models on their own workloads, while users should distinguish official services from unofficial resellers. Those steps do not establish what happened; they address the procurement and access questions raised by the report.
Key Questions
What does The Information allege?
It reported that DeepSeek and Moonshot AI relied heavily on GPT-6 Astra outputs to train recent models. The report cites people with direct knowledge, and Reuters summarized the allegation; the companies had not confirmed it in Reuters’ account.
Which models are named?
The report names Moonshot’s Kimi K3.1 Pro and DeepSeek-V4.1-Flash-1017. The source material gives release dates of Oct. 23 and Oct. 17, 2026, respectively.
Is distillation itself improper?
No. Distillation is a common AI training technique, including when a lab uses its own larger model to train a smaller one. The dispute described in the report concerns alleged use of a competitor’s outputs and whether access or use breached the provider’s terms.
Has OpenAI confirmed the allegation?
No confirmation is included in the Reuters account described here. Reuters reported that OpenAI, DeepSeek and Moonshot did not immediately respond. SBS separately reported an OpenAI spokesperson’s general statement that the company takes the issue seriously, but that does not confirm the specific allegation.
Does the report show the models copied Astra?
The public information described here does not establish that. It gives no figures, detailed method or account of how outputs were obtained, and does not quantify any effect on the models. The claim remains unverified in the available record.
Source: ThorstenMeyerAI.com
