🔍 Read the full analysis: What We Know About China’s New AI Models And GPT-6 Astra Distillation on ThorstenMeyerAI.com
Get hardware and tech essentials 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 that DeepSeek and Moonshot AI relied heavily on OpenAI’s GPT-6 Astra outputs to train recently released models. Reuters said the three companies did not respond to requests for comment; public reporting does not specify the scale or method, and the allegation remains unconfirmed.
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 models released this month. Reuters summarized the allegation, but said OpenAI, DeepSeek and Moonshot did not immediately respond; the public accounts do not establish how many outputs were used or how they were obtained.
The models named in the reporting are Moonshot’s Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, released October 17. The Information’s public article opening described Astra as a powerful model released in early September and said Chinese AI labs were among a growing customer base seeking to extract its capabilities. The full article is paywalled, and the public material provides no figures for the volume of outputs allegedly used.
Reuters reported the central claim as the labs having “relied heavily” on Astra outputs to train their latest models. Its account said the companies did not immediately respond to requests for comment. That leaves the allegation without a public confirmation or denial from the companies named. Moonshot’s Kimi launch post attributed gains to larger post-training compute and large-scale reinforcement learning; DeepSeek’s update note, as described in the source material, did not discuss training data.
South Korean broadcaster SBS added that sources told it 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 to work on it. Those details have not been independently checked against The Information’s full report in the supplied material, so they remain attributed to SBS rather than established findings.
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 AI Model Training
The allegation matters because distillation can transfer a model’s capabilities into a cheaper or more compact system. A lab can use a larger model’s answers as training examples for a smaller “student” model. The practice is commonly used within AI development; the dispute is whether one company’s model was used at scale by competitors and whether access was obtained in ways that breached terms or involved deception.
OpenAI’s terms, according to the source material, prohibit using its outputs to build competing models. If the reported conduct occurred, that could raise contract and access questions, alongside questions about how the outputs were collected. The available reporting does not establish a legal violation, and whether any conduct broke a law would depend on evidence and applicable law. This is not a legal finding.
The source material gives a price and performance comparison, but it does not show that the reported training caused those results. 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 at peak hours. On the same independent index, their scores are reported as 57, 46 and 32, respectively. Those figures suggest lower listed prices alongside lower scores on that index—not equivalent performance. The index’s name and testing details are not provided here.
For businesses choosing models, the report is a reminder to ask vendors about training-data provenance and access controls, test systems on their own workloads, and have a fallback option. The source advises users to avoid cheap, unofficial API resellers. It does not report a change to the named models’ availability or say that ordinary users’ apps have changed.
Products Worth Considering
Apple MacBook Pro 16.2-inch with M5 Pro, 18-core CPU, 20-core GPU, 48GB Memory, 1TB SSD, Wi-Fi 7, Space Black
Apple MacBook Pro 14‑inch (2021) with M1 Pro, 16 GB RAM, 512 GB SSD – Space Gray
Logitech MX Mechanical Wireless Illuminated Keyboard Tactile (Graphite)
As an affiliate, we earn on qualifying purchases.
Distillation and Earlier Allegations
Distillation is a standard training technique: a teacher model generates answers to prompts, and a student model is trained to imitate them. It can make a model less expensive to run. The U.S. government advisory described in the source material calls distillation “a legitimate and useful technique in AI research.” The concern raised in this case is not the technique by itself, but the alleged use of a rival’s outputs and the means used to obtain them.
The source material also describes earlier allegations involving access to competing models. It says OpenAI linked a July campaign to extract hidden reasoning from its models to people associated with Moonshot AI, and that U.S. security agencies named DeepSeek and Moonshot in an advisory. China’s commerce ministry called the accusations baseless. These statements provide background, but they do not independently prove the new claim about Astra training.
There is also a distinction between ordinary answers and hidden reasoning traces. Providers may restrict access to detailed reasoning, which can be more useful for training than short final answers. The supplied material cites an Anthropic experiment in which a model trained on full reasoning transcripts scored 80.4% on SWE-bench Verified, compared with 71.8% when trained on summarized transcripts. That is Anthropic’s own experiment and does not establish what data DeepSeek or Moonshot used.
“DeepSeek and Moonshot “relied heavily on outputs” from GPT-6 Astra to train their latest models.”
— The Information, as summarized by Reuters
Evidence Behind the Astra Claim
The allegation remains unverified publicly. The Information cited people with direct knowledge, but its detailed account is behind a paywall. The supplied reporting does not provide the number of prompts or outputs involved, the dates of any activity, the collection method, or evidence that hidden reasoning was obtained.
It is also unclear whether the labs used direct accounts, resold access, or another route. The source material says fake or resold accounts are part of the broader concern around industrial-scale distillation, but does not establish that either lab used them in this case. No public technical analysis, training records or independent audit is cited to substantiate the specific claim.
The named companies had not publicly confirmed or denied the allegation in Reuters’ account. It is also unclear whether OpenAI has taken any action, whether the alleged conduct caused measurable changes in either model, or whether the matter has led to a legal or regulatory proceeding.
Company Responses and Verification
The next developments to watch are direct responses from OpenAI, DeepSeek and Moonshot, along with any further reporting that supplies evidence about the alleged data collection and training. The source material does not identify a scheduled investigation, hearing or company announcement, so no formal next milestone can be confirmed.
For now, the distinction is between the reported allegation and verified public facts: the models were released on the dates stated, The Information made the claim, and Reuters reported that the companies did not immediately respond. More detail on methods, scale and any consequences would be needed to assess the claim beyond that account.
Key Questions
Which models are named in the report?
The report names Moonshot AI’s Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, released October 17.
Has OpenAI confirmed that the models were trained on Astra outputs?
No public confirmation is described in the source material. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment.
Is distillation itself improper?
Not inherently. Distillation is a common way to train a smaller model using a larger model’s outputs. The reported dispute concerns the alleged use of a competitor’s model and whether access or use breached the provider’s terms or involved deception.
Does the report show how the outputs were collected?
No. The public details described here do not specify the volume, method or account access route, and do not establish that hidden reasoning was obtained.
Have users’ apps or model access changed?
The source material reports no change to user-facing apps or model availability. It advises caution with cheap, unofficial API resellers; AI services and their outputs can vary, and users should check providers’ terms and reliability.
Source: ThorstenMeyerAI.com
