🔍 Read the full analysis: Does China’s New AI Model Report Establish A GPT-6 Astra Connection? 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 reported on Oct. 28 that DeepSeek and Moonshot AI relied heavily on OpenAI’s GPT-6 Astra outputs to train recent models. Reuters said the companies involved did not respond to its requests for comment; the public record does not disclose how the outputs were obtained or how much they contributed.
The Information reported on Oct. 28 that Chinese AI labs DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra to train recent models. Reuters summarized the allegation, naming Moonshot’s Kimi K3.1 Pro and DeepSeek’s V4.1 Flash update; none of the three companies had publicly confirmed or denied it in the reporting cited.
The Information’s report, by Jing Yang, Juro Osawa and Qianer Liu, was published at 6 p.m. Pacific time on Oct. 28. Its public opening described Astra as OpenAI’s most powerful released model and said Chinese AI labs were among a growing customer base seeking to extract its capabilities. The full story is paywalled, and its public material does not provide figures for the volume of outputs or explain the methods used to obtain them.
Reuters reported that the models named were Kimi K3.1 Pro, released Oct. 23, and DeepSeek-V4.1-Flash-1017, launched Oct. 17. 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 larger post-training compute” and large-scale reinforcement learning; DeepSeek’s update note did not describe its training data, according to the source material.
South Korea’s SBS offered further details on Oct. 29, saying sources described a shift toward OpenAI models after Anthropic tightened defenses against extraction. SBS also quoted an OpenAI spokesperson as saying the company takes the issue very seriously and continues working on it. Those details have not been independently checked against The Information’s full report, which is not publicly available.
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 Sources Matter
If substantiated, the allegation would put model-output sourcing and commercial terms at the center of a dispute over how AI developers compete. Distillation—training one model on answers generated by another—can reduce the cost and time needed to build a capable system. But using a rival’s outputs at scale may violate that provider’s terms, depending on the terms and how the access was obtained.
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. It also reports scores of 57, 46 and 32, respectively, on the same independent index, without identifying the index or its methodology in the supplied material. These figures do not establish that training on Astra caused any performance difference.
For businesses choosing AI systems, the report is a reason to ask vendors about training-data provenance and to test models on their own tasks. It does not show that either named model is unusable or that users’ apps have changed. The source advises users to avoid cheap, unofficial API resellers, but does not establish that any particular reseller supplied access for the alleged training.
Products Worth Considering
Logitech MX Keys S Wireless Keyboard, Graphite
Samsung 27-Inch Odyssey QD-OLED G8 4K 240Hz Gaming Monitor
Anker SOLIX S2000 Portable Power Station, 2,010Wh, 1,500W
As an affiliate, we earn on qualifying purchases.
Distillation and Earlier Allegations
Distillation is a standard training technique: a smaller “student” model is trained using outputs from a larger “teacher” model. Developers commonly apply it to their own models. A U.S. government advisory cited in the source material describes distillation as “a legitimate and useful technique in AI research.” The controversy is not the technique by itself, but whether a rival’s outputs were collected at scale in violation of terms or through deceptive access.
The report follows earlier public accusations. In September, OpenAI linked what it called a “core cluster” of a July campaign to extract hidden reasoning from its models to individuals associated with Moonshot AI, according to the source material. A U.S. security advisory also named DeepSeek and Moonshot, while China’s commerce ministry called the accusations baseless. These are separate claims and responses; they do not independently prove the new report’s account of training.
OpenAI’s terms, as summarized in the source material, prohibit using its outputs to build competing models. Whether the alleged conduct breached those terms would depend on evidence about the access, use and applicable agreements. The supplied material characterizes the issue primarily as a contractual question, with potential concerns about fraud if access was obtained through fake or resold accounts; it provides no finding by a court or regulator.
“Substantially larger post-training compute”
— Moonshot AI, in its Kimi K3.1 Pro launch post
Evidence Behind the Astra Claim
The central allegation remains unconfirmed in public statements from OpenAI, DeepSeek or Moonshot, according to the cited reporting. The available account does not identify the specific prompts, quantify the outputs, explain whether hidden reasoning was obtained, or establish how access was arranged. Without those details, it is not possible to assess the scale of any copying or how much it shaped either model.
The report relies on unnamed sources, and its full text is paywalled. SBS’s account adds claims about earlier reliance on Anthropic models and an OpenAI response, but the source material says those details could not be checked against the original. It is also unclear whether any terms were breached, whether accounts were misrepresented or resold, and whether OpenAI has opened a formal legal or regulatory process. No legal finding is described.
Company Responses and Evidence to Watch
The next developments to watch are direct responses from the three companies, any additional reporting that sets out the evidence, and whether OpenAI describes enforcement or changes to access controls. Public technical details about the models’ training data or methods could help test the allegation, but none are provided in the source material.
For now, the report should be treated as an attributed allegation rather than an established account of how Kimi K3.1 Pro or DeepSeek’s V4.1 Flash update were trained. The supplied material reports no immediate change to the consumer apps. Companies evaluating either model can continue testing them while asking providers about data provenance, service terms and support for independent due diligence.
Key Questions
What does the report allege?
The Information said DeepSeek and Moonshot AI relied heavily on GPT-6 Astra outputs to train recent models. The companies have not publicly confirmed the claim in the reporting cited.
Which models are named?
The report names Moonshot’s Kimi K3.1 Pro, released Oct. 23, and DeepSeek-V4.1-Flash-1017, launched Oct. 17, according to Reuters’ summary.
Is distillation itself improper?
No conclusion follows from the technique alone. Distillation is commonly used to train smaller models. The dispute concerns whether a rival’s outputs were used at scale in a way that violated terms or involved deceptive access.
Has OpenAI confirmed that its terms were violated?
No such finding is reported. The source material says OpenAI’s terms prohibit using its outputs to build competing models, but the public account does not establish the access method, the applicable agreement or a legal outcome.
Does this change what users should do now?
The source material reports no immediate change to the apps. Businesses can ask vendors about training-data provenance and test models for their own uses; users are advised in the source to avoid cheap, unofficial API resellers.
Source: ThorstenMeyerAI.com
