🔍 Read the full analysis: The Evidence And The Gaps In The Report On China’s AI Models And GPT-6 Astra 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 outputs from OpenAI’s GPT-6 Astra to train recent models, citing people with direct knowledge. The companies did not respond to Reuters, and public details about the alleged data, methods and scale are unavailable. Distillation is a standard AI technique; whether it violated terms or involved deceptive access remains unproven.
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 recent models, a claim that could raise questions about model access and the use of rival systems’ outputs. Reuters summarized the report, which cited people with direct knowledge; none of the three companies had confirmed or denied the allegation in responses reported by Reuters.
The models named in the report are Moonshot’s Kimi K3.1 Pro, released Oct. 23, and DeepSeek-V4.1-Flash-1017, released Oct. 17. Reuters said The Information reported that both models “benefitted from distilling” Astra. The publicly available opening of The Information’s story describes Chinese AI labs as a growing customer base for Astra, but the rest of the article is paywalled, limiting independent examination of its evidence.
The report, as summarized publicly, does not provide the number of Astra outputs involved, how the outputs were obtained, or what proportion of either model’s training they represented. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment. Moonshot’s launch post attributed Kimi K3.1 Pro’s gains to “substantially larger post-training compute” and large-scale reinforcement learning. DeepSeek’s update note, according to the source material, did not describe its training data.
A separate account by South Korea’s SBS on Oct. 29 said Chinese labs had previously focused on Anthropic’s Claude and shifted toward OpenAI after Anthropic strengthened defenses. SBS also reported that an OpenAI spokesperson said the company takes the issue seriously and continues working on it. Those details could not be checked against the paywalled report, so they are secondary reporting rather than independently verified evidence.
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 of Model Output Use
The allegation matters because distillation can transfer capabilities from a larger model to a cheaper one. If a rival’s outputs are gathered at scale, the resulting model may reproduce some of its performance without the same cost of developing the original system. That possibility makes the source and terms of training data relevant to AI companies, customers and regulators.
But the report does not establish that the alleged activity breached a contract, involved fraud, or produced a particular level of capability. The source material says OpenAI’s terms prohibit using its outputs to build competing models; whether those terms applied to the activity described, and whether they were violated, would depend on facts not made public. The allegation is not a legal finding.
For businesses choosing models, the episode is a reminder to ask vendors about training-data provenance and contractual assurances, while testing systems on their own workloads. It does not by itself show that either named model is unsafe or unusable. The report also says nothing changes immediately for users of the products; the source advises caution with unofficial, low-cost API resellers, whose access arrangements may be unclear.
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
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
Apple MacBook Pro (14.2-inch, 2025) – M5 chip, 16GB Unified Memory, 1TB SSD, Space Black
As an affiliate, we earn on qualifying purchases.
How Distillation Became the Issue
Distillation is a common training approach: a “student” model learns from answers generated by a larger “teacher” model. The source material describes it as widely used by labs to create smaller, less expensive versions of their own systems. A U.S. government advisory cited in that material calls distillation “a legitimate and useful technique in AI research.” The controversy concerns whose outputs are used, at what scale, and how access is obtained, not the technique alone.
The report also lands amid earlier accusations. According to the source material, OpenAI said in September that it had linked a “core cluster” of a July effort to extract hidden reasoning from its models to people associated with Moonshot AI. It also says U.S. security agencies named DeepSeek and Moonshot in an advisory, while China’s commerce ministry called the accusations baseless. These are separate claims and responses; they do not independently prove the new allegation about Astra or either model’s training.
One reason model outputs may be valuable is that detailed reasoning traces can provide more information than final answers alone. The source cites an 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 experiment, not evidence about Astra, DeepSeek or Moonshot.
“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 Astra Allegation
The public record does not disclose the underlying evidence, output volume, training method or access route. The Information cited people with direct knowledge, but the public opening does not identify them, and the rest of its report is paywalled. The claim therefore remains a reported allegation rather than a publicly documented finding.
It is also unclear whether outputs were gathered through ordinary accounts, resellers, automated access, or other means; whether hidden reasoning was involved; and how much any such data affected the released models. The companies’ lack of an immediate response to Reuters is not confirmation or denial. No public technical analysis presented in the source material establishes that the models’ behavior or benchmarks demonstrate Astra-derived training.
Responses and Evidence to Watch
The next developments to watch are direct responses from OpenAI, DeepSeek and Moonshot, along with any fuller publication of the reporting’s evidence. Technical documentation, account or access records, and independent analysis could clarify whether Astra outputs were used and at what scale. Contractual or regulatory consequences would depend on facts and applicable rules not established in the current public record.
Until further evidence emerges, readers should distinguish the reported claim from confirmed information: the named models and release dates are public, while the alleged training relationship and its methods remain unverified. Companies evaluating the models can conduct their own task-specific tests and seek clarity from vendors on training-data provenance.
Key Questions
Which Chinese AI models does the report name?
The report names Moonshot AI’s Kimi K3.1 Pro, released Oct. 23, and DeepSeek-V4.1-Flash-1017, released Oct. 17.
Has OpenAI confirmed that its outputs were used?
No. Reuters reported that OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment. The allegation has not been publicly confirmed by the companies.
Is distillation itself improper?
Not necessarily. Distillation is a standard way to train a smaller model using outputs from a larger one. The dispute in this report concerns use of a rival’s outputs, scale and access methods, as well as any applicable terms.
Does the report prove the models broke OpenAI’s terms?
No. The source material says OpenAI’s terms prohibit using its outputs to build competing models, but the public report does not establish how the outputs were obtained or whether a contract was breached. That remains unresolved.
Source: ThorstenMeyerAI.com
