🔍 Read the full analysis: The Report On China’s New AI Models: Claims About GPT-6 Astra, And Their Limits 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 October 28 that DeepSeek and Moonshot AI relied heavily on GPT-6 Astra outputs to train recent models. Reuters repeated the claim, but the evidence is not public, the companies did not respond to Reuters, and the report gives no figures or details about how the outputs were obtained.
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 recent models, including DeepSeek-V4.1-Flash-1017 and Kimi K3.1 Pro. Reuters summarized the report, but the underlying evidence remains largely unavailable to the public, and none of the three companies responded to Reuters before its account was published.
The Information’s report, by Jing Yang, Juro Osawa and Qianer Liu, cited people with direct knowledge. Its public opening said Astra had drawn strong demand since its release in early September and that Chinese AI labs were among customers seeking to extract its capabilities. The rest of the story is behind a paywall. Reuters reported the central allegation: that DeepSeek and Moonshot’s latest models benefited from distilling Astra.
The models identified were Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, launched October 17. The report does not give a volume of Astra outputs used, explain the alleged training process, or say how the labs obtained access. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to requests for comment.
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 its training data, according to the source material. Neither public description independently confirms or disproves The Information’s account.
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 allegation matters because distillation can transfer capabilities from a high-cost model into a cheaper one. In the ordinary process, a teacher model produces answers to many prompts and a student model trains on them. Labs use the technique on their own systems; the dispute here is whether the labs used a competitor’s outputs at scale and, if so, whether they obtained them in ways that breached contractual rules or involved deception.
OpenAI’s terms prohibit using its outputs to build competing models, according to the source material. If the reported conduct occurred, it could become a contract and access-control dispute, as well as a question about how model providers can protect their services. The available reporting does not establish that any law was broken, and the legal implications depend on facts that are not public.
Price and performance comparisons also help explain the commercial stakes, but do not establish how a model was trained. The source lists Astra at $10 per million input tokens and $50 per million output tokens, compared with 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, the models scored 57, 46 and 32 respectively. Those figures are reported comparisons, not proof that lower prices or scores resulted from distillation.
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
FLEXISPOT Standing Desk with Monitor Riser & Cable Management
FLEXISPOT 42-Inch Electric Standing Desk Converter
As an affiliate, we earn on qualifying purchases.
Earlier Concerns About Model Extraction
The report follows earlier public concern about attempts to extract capabilities from AI systems. The source material says that in September, OpenAI linked a “core cluster” of a July campaign targeting its models’ hidden reasoning to individuals associated with Moonshot AI. It also says U.S. security agencies named DeepSeek and Moonshot in an advisory. China’s commerce ministry called the accusations baseless. These are separate claims and responses; they do not independently verify the new report about Astra.
Distillation is not inherently improper. A U.S. government advisory described it as a legitimate and useful AI research technique, according to the source. The concern in this case is whether a rival’s model was used, at what scale, and whether access complied with the provider’s terms. The source also relays, through South Korean broadcaster SBS, an account that labs shifted toward OpenAI after Anthropic tightened defenses, along with a statement attributed to an OpenAI spokesperson. Those details could not be checked against The Information’s paywalled report and should be treated as secondhand.
“Astra, “the most powerful model the company has released,” has met strong demand since its release in early September.”
— The Information, in the public opening of its October 28 report
Evidence Behind the Astra Allegation
The key evidence has not been made public in the material available here. The Information cited unnamed sources, while Reuters summarized its central claim; the full report is paywalled. There are no disclosed counts of prompts or outputs, no technical analysis of the training data, and no public account of whether access came through ordinary accounts, resellers or another route.
None of the three companies had responded to Reuters at publication time, so their positions on this specific report are not established. The source says the U.S. government advisory named the labs and that China’s commerce ministry rejected the accusations, but neither fact proves or refutes the claim about Astra. It also remains unclear how much any alleged use contributed to model performance, and whether the models’ published training explanations account for their development.
Company Responses and Further Evidence
The next developments to watch are responses from OpenAI, DeepSeek and Moonshot, along with any additional reporting that makes the evidence, scale or access methods clearer. The source material gives no timetable for a company statement, investigation or legal action, so none should be assumed.
For organizations evaluating AI systems, the immediate practical point is to check model provenance and test performance on their own tasks rather than infer quality or training history from price alone. The source also advises keeping a fallback model and avoiding unofficial, unusually cheap API resellers. For individual users, it reports no announced change to the apps as a result of the story. These steps do not resolve the allegation; they address uncertainty while the facts remain incomplete.
Key Questions
What did The Information report?
It reported that DeepSeek and Moonshot AI relied heavily on GPT-6 Astra outputs to train recent models. Reuters summarized the claim, citing The Information’s reporting.
Which models were named?
The report named Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, launched October 17.
Has the claim been confirmed by the companies?
No public confirmation is included in the source material. OpenAI, DeepSeek and Moonshot did not immediately respond to Reuters, and the report’s underlying evidence is not publicly detailed.
Is AI distillation itself improper?
Not necessarily. Distillation is a standard training technique and has been described as useful in AI research. The disputed questions are whether a lab used a competitor’s outputs, whether the use breached service terms, and how access was obtained.
What is still unknown?
The public record does not specify how many outputs were used, how they were obtained, or how much they affected model performance. The companies’ responses to the specific allegation are also not included.
Source: ThorstenMeyerAI.com
