🔍 Read the full analysis: China’s AI Models, GPT-6 Astra, And The Distillation Evidence Explained 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 October 28 that DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra when training recent models. Reuters said the three companies did not immediately respond, and the public reporting does not specify how the outputs were obtained or how much training depended on them.
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. Reuters repeated the allegation, but its account said OpenAI, DeepSeek and Moonshot did not immediately respond; the public reporting does not establish how the outputs were obtained or how much they contributed to training.
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 opening described Astra as OpenAI’s most powerful model and said Chinese AI labs were among a growing customer base seeking to extract its capabilities. The rest of that report is paywalled. Reuters summarized its central allegation as the labs having relied heavily on Astra’s outputs.
The account rests on unnamed sources, according to the material available publicly. It contains no figures for the volume of outputs, no description of the collection method and no independent technical evidence establishing what training data was used. Moonshot’s launch post attributed Kimi K3.1 Pro’s gains to more post-training compute and large-scale reinforcement learning. DeepSeek’s update note, as described in the source material, did not detail the model’s training data.
Reuters reported that the three companies did not immediately respond to its requests for comment. Korean broadcaster SBS separately summarized the report and quoted an OpenAI spokesperson saying the company takes the issue seriously and continues to work on it. SBS also reported claims about labs’ earlier use of another provider’s models, but those details could not be checked against The Information’s paywalled account and are not independently established here.
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
If the reported use is substantiated, the dispute would put a spotlight on a central question in AI competition: when does learning from a model’s answers become unauthorized copying? Distillation can reduce the cost of building a capable model, but using a rival’s paid outputs at scale may conflict with that provider’s terms. If access involved deceptive accounts or resellers, that could raise separate concerns; the public record provided so far does not establish that happened in this case.
The source material gives a price and benchmark comparison, but it should not be mistaken for evidence of how the models were trained. It 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 at $0.30 and $1.20 during peak hours. On the same independent index, the reported scores were 57, 46 and 32, respectively. The figures suggest different price-performance positions, not equivalent capability or proof of copied training.
For businesses, the report is a reminder to ask vendors about training-data provenance, contractual assurances and model evaluation. Companies can also test models on their own tasks and retain alternatives if vendor access or reliability changes. For individual users, the source says there is no immediate change to the named apps; it advises caution with unofficial, unusually cheap API resellers. That is practical guidance, not evidence that any particular reseller was involved.
Products Worth Considering
Logitech MX Mechanical Wireless Illuminated Keyboard Tactile (Graphite)
Anker SOLIX S2000 Portable Power Station, 2,010Wh, 1,500W
Jackery HomePower 3000 Portable Power Station with Two 200W Solar Panels
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 responses from a larger teacher model. AI labs commonly apply it to their own systems to create more efficient models. A U.S. government advisory cited in the source material describes distillation as a legitimate and useful research technique. The controversy is not the technique by itself, but whether a lab used another company’s outputs at scale in ways barred by its terms or obtained through improper access.
The source material says OpenAI’s terms prohibit using its outputs to build competing models. Whether that restriction was breached would depend on the relevant terms, the conduct and evidence, and potentially legal interpretation; this report does not establish a violation. The material also describes a broader backdrop: OpenAI previously connected a July campaign seeking model reasoning to people it associated with Moonshot AI, while U.S. security agencies named both labs in an advisory. China’s commerce ministry called those accusations baseless. Those claims and denials are separate from the new report and do not prove its allegation.
Other companies have raised similar concerns about large-scale extraction. The source cites Anthropic research comparing training on full reasoning transcripts with summarized transcripts; in its experiment, the model trained on full transcripts scored 80.4% on SWE-bench Verified, versus 71.8% for the summarized version. That company-run result illustrates why detailed outputs may be valuable, but it does not verify that DeepSeek or Moonshot obtained or used Astra reasoning transcripts.
“DeepSeek and Moonshot relied heavily on outputs from OpenAI’s GPT-6 Astra model to train their latest models.”
— The Information, as summarized by Reuters
Evidence Behind the Training Claim
The allegation has not been publicly substantiated with training records, output counts, account details or a technical analysis linking Astra responses to either model. The Information’s full report is behind a paywall, and the public summary relies on unnamed sources. The companies’ lack of an immediate response to Reuters is not confirmation or denial.
It is also unclear whether any Astra outputs were obtained through ordinary paid access, accounts controlled by others, resellers or another route. The reporting as presented does not show whether hidden reasoning was involved, whether the alleged use was authorized, or how much any such data affected model performance. Those details matter to both the technical claim and any question about contractual or legal responsibility.
Responses and Verification Ahead
The next developments to watch are direct responses from OpenAI, DeepSeek and Moonshot, plus any further reporting that supplies evidence beyond anonymous sourcing. Relevant details would include the volume and type of outputs, how access was obtained, and whether the companies’ terms were breached. Public training documentation or independent technical analysis could help test the claims, though neither is promised.
For now, the report is a significant allegation rather than a settled account of what happened. Companies evaluating these models can review vendor assurances and data-provenance practices, while users should distinguish official services from third-party API offers. No enforcement action, formal finding or confirmed change to the models’ availability is identified in the material available at publication.
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, naming DeepSeek-V4.1-Flash-1017 and Moonshot’s Kimi K3.1 Pro. The claim is based on unnamed sources in the public account.
Have the companies confirmed the report?
No. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment. That does not establish whether the allegation is true or false.
Is distillation itself improper?
No. Distillation is a common training method, including for making smaller versions of a lab’s own models. Concerns arise if a provider’s outputs are used in ways barred by its terms or obtained through improper access; the current report does not establish those details.
What evidence is publicly available?
The public material provides the reported allegation and the names of the models, but no output counts, access method or training-data audit. The full Information article is paywalled, and the named companies have not publicly confirmed the claim in the cited Reuters account.
Source: ThorstenMeyerAI.com
