🔍 Read the full analysis: China’s New AI Models: Weighing The GPT-6 Astra Distillation Report on ThorstenMeyerAI.com
Get the latest gadgets 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, citing people with direct knowledge, reported that DeepSeek and Moonshot AI relied heavily on OpenAI’s GPT-6 Astra outputs to train recent models. Reuters reported that none of the three companies responded to its requests for comment; the public record does not establish how the outputs were obtained or how much they shaped the models.
The Information reported on October 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 compliance with OpenAI’s terms. Reuters summarized the report within hours, but the companies have not publicly confirmed the account, and details about the training process remain unavailable.
The models identified 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 Chinese AI labs as a growing customer group seeking to extract capabilities from Astra. Reuters reported the central allegation that the two labs relied heavily on Astra outputs. The underlying Information article is paywalled, and its claims are attributed to people with direct knowledge rather than named sources.
Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to requests for comment. The source material also cites a report by South Korean broadcaster SBS, which said OpenAI characterized the issue as serious and said it was continuing to work on it. That wording has not been independently checked against the original Information article in the supplied material. Neither model’s launch materials publicly describe training on Astra: Moonshot credited Kimi’s gains to more post-training compute and reinforcement learning, while DeepSeek’s update note did not discuss its training data.
The report provides no figures for the volume of outputs, no detailed account of how they were accessed, and no evidence in the public material establishing how much the alleged training contributed to either model’s performance. Those gaps limit what can be concluded from the allegation alone. The fact that a model can imitate another model’s answers is not, by itself, proof that a particular company did so or violated a rule.
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 Provenance Matters
If the report is substantiated, it could intensify scrutiny of how AI developers obtain training material and whether commercial model outputs can be used to create competing systems. OpenAI’s terms prohibit using its outputs to build competing models, according to the source material. Whether any conduct described in the report breached those terms would depend on facts that have not been made public, including the access method, the terms applying at the time, and the scale and purpose of the activity.
The allegation also matters to organizations choosing models. The source’s cited prices put 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 in the source, the models scored 57, 46 and 32, respectively. Those figures are not proof of equivalent capability or a direct measure of how any model was trained; they show that lower listed prices and different scores can influence purchasing decisions. Buyers may want to ask vendors about data provenance and maintain alternatives, while testing models on their own tasks.
For users, the report does not establish a change to the models’ current availability or behavior. The immediate concern is less about a confirmed disruption than about trust, contractual boundaries and the origin of model capabilities. The source advises caution with cheap, unofficial API resellers, but it does not report that a specific reseller was involved in this case.
Products Worth Considering
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
Bose Lifestyle Ultra Wireless Home Speaker
As an affiliate, we earn on qualifying purchases.
Distillation and the Earlier Warnings
Distillation is a training method in which a smaller “student” model learns to produce answers similar to a larger “teacher” model. AI labs commonly use it to make their own systems cheaper or more efficient. The U.S. government advisory cited in the source calls distillation a legitimate and useful research technique. The contested issue is whether a developer used another company’s model outputs at scale in ways restricted by the provider’s terms or obtained access through improper means.
The allegation follows earlier public warnings, but those warnings do not prove the new report’s specific claims. The source says OpenAI linked a July campaign seeking to extract hidden reasoning from its models to individuals associated with Moonshot AI, and that U.S. security agencies named DeepSeek and Moonshot in an advisory. It also says China’s commerce ministry called the accusations baseless. These are attributed positions and allegations, not an independent finding that the labs used Astra to train the two models named in the current report.
The source further describes an earlier Anthropic account involving large numbers of exchanges with Claude, but the supplied excerpt ends before giving full details. It also cites Anthropic’s own experiment suggesting that training on full reasoning transcripts can produce different results from training on summaries. That offers context for why access to detailed model outputs may matter, but it does not establish that Astra’s hidden reasoning was obtained or used in this case.
“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
What the Public Record Cannot Show
The available reporting does not disclose how many Astra outputs were allegedly used, which prompts or accounts were involved, whether hidden reasoning was accessed, or what role any such data played in either model’s training. It is also unclear whether the access, if it occurred, was direct, through a reseller, or by another route. Without those details, the public cannot assess the scale of the alleged activity or determine whether it violated a contract or law.
The allegation remains unconfirmed by the named companies in the supplied material. Reuters reported that it received no immediate responses, which is not the same as a denial. The article also does not present technical evidence, training records, or an independent analysis linking Astra outputs to the released models. The distinction between reported claims, earlier warnings and verified facts remains central.
Company Responses and Evidence
The next developments to watch are public responses from OpenAI, DeepSeek and Moonshot, along with any additional reporting that identifies the methods, scale or documentation behind the allegation. Technical evidence or disclosures from the companies would be needed to clarify whether and how Astra outputs entered the models’ training processes.
For now, organizations evaluating these systems can review contractual terms, ask providers about training-data provenance and test models against their own requirements. Those steps are prudent responses to uncertainty, not evidence that the reported conduct has been proven. The underlying claims may become clearer if the companies, regulators or the reporters provide further documentation.
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, citing people with direct knowledge. The public account does not include detailed methods or quantities.
Which models were named?
The report named Moonshot’s Kimi K3.1 Pro and DeepSeek-V4.1-Flash-1017. The source says they were released October 23 and October 17, respectively.
Have the companies confirmed the allegation?
No confirmation is included in the supplied reporting. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment; that does not establish that they denied the claim.
Is distillation itself improper?
Distillation is a common training technique and is described in the cited U.S. advisory as useful in AI research. The dispute concerns whether a rival’s outputs were used in a way prohibited by its terms, and whether access was obtained improperly. The available report does not establish those facts.
What is still unknown?
The public material does not specify the amount or type of Astra output allegedly used, how it was obtained, or how much it affected either model. It also does not provide evidence resolving whether any contract or law was breached.
Source: ThorstenMeyerAI.com
