🔍 Read the full analysis: Did China’s Latest AI Models Learn From GPT-6 Astra? Here’s What’s Reported on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
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. Reuters relayed the allegation, but the companies did not respond to its requests for comment, and public reporting has not disclosed the method or scale. The claim concerns possible use of a rival model’s outputs, not whether AI distillation itself is legitimate.
The Information reported on Oct. 28 that Chinese AI labs DeepSeek and Moonshot AI relied heavily on responses from OpenAI’s GPT-6 Astra to train recently released models. Reuters relayed the report within hours, but said OpenAI, DeepSeek and Moonshot did not immediately respond to requests for comment, leaving the central allegation unconfirmed publicly.
The models named are Moonshot’s Kimi K3.1 Pro, released Oct. 23, and DeepSeek-V4.1-Flash-1017, released Oct. 17, according to Reuters’ summary of The Information’s reporting. The Information’s public article opening described Astra as a strong draw for Chinese AI labs seeking to extract capabilities from it; the rest of its report is behind a paywall. Its account cites people with direct knowledge, but the public material does not identify them.
Details about the alleged training are not public: the reporting available here gives no figures for how many Astra responses were used, which prompts were sent, how the responses entered training, or how much the models’ performance depended on them. Moonshot’s launch post credited 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. Those company statements do not independently resolve the allegation.
Reuters reported that none of the three companies had responded by the time it published its account. A separate summary by South Korean broadcaster SBS attributed a statement to an OpenAI spokesperson saying the company takes the issue very seriously and continues working on it. That statement is reported by SBS; it is not a direct response from OpenAI included in Reuters’ 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
Why Model-Output Access Matters
The report matters because it puts a familiar AI training technique into a dispute over who controls access to model outputs. Distillation—training a smaller model to imitate a larger one’s responses—is widely used, including by labs training smaller versions of their own systems. The contested issue is whether a company used a rival’s model at scale, and whether it did so in a way that violated the provider’s terms or involved improper access.
If the allegation is substantiated, it could raise questions for AI developers about contract enforcement, account controls and training-data provenance. Customers evaluating models may also want to know how systems were trained and whether their providers can explain the source of their data. Those are practical diligence questions, not proof that any particular model or dataset is unlawful.
Users should not infer from the report that the named models are unusable or that their apps have changed. The source material says no immediate user-facing change has been announced. It does advise caution around cheap, unofficial API resellers, though it provides no evidence that a specific reseller was involved in the reported training.
Popular right now
Roborock F25 Ultra Steam Wet Dry Vacuum Mop
Ninja Flip Toaster Oven and Air Fryer, 8-in-1, Stainless Steel, SP151
Ninja Hot & Iced XL Coffee Maker with Rapid Cold Brew, Black, CM371
As an affiliate, we earn on qualifying purchases.
Distillation and Earlier Allegations
In distillation, a teacher model generates answers to many prompts, and a student model is trained to reproduce or learn from those answers. A U.S. government advisory cited in the source material describes distillation as a legitimate and useful research technique. The method itself does not establish wrongdoing; the dispute depends on who supplied the outputs, how access was obtained and what the provider’s rules say.
The latest report follows earlier public allegations involving access to competitors’ models. In September, OpenAI said a “core cluster” from a July campaign to extract hidden reasoning from its models was linked to individuals associated with Moonshot AI. U.S. security agencies also named DeepSeek and Moonshot in an advisory, according to the supplied account. China’s commerce ministry called the accusations baseless. These earlier claims provide context but do not establish that the newly named models were trained on Astra outputs.
The source material also describes an earlier Anthropic allegation involving large volumes of exchanges with Claude. Those claims and the Astra report are separate matters. The public information summarized here does not establish a link between them or verify how any company obtained the outputs at issue.
Evidence Behind the Astra Claim
The central claim remains unconfirmed in public statements from OpenAI, DeepSeek or Moonshot, based on the responses reported by Reuters. The public account relies on unnamed sources and does not provide the underlying records, such as prompts, account data, training logs or a technical comparison demonstrating how Astra outputs shaped either model.
It is also unclear whether outputs were obtained through ordinary paid access, resold or shared accounts, or another route. The source material says OpenAI’s terms prohibit using its outputs to build competing models, but whether the reported conduct occurred and whether it violated those terms would depend on facts not yet disclosed. No legal finding is reported.
Further, the available reporting does not quantify the alleged contribution or establish that the models’ capabilities came primarily from Astra rather than their developers’ own training methods. The claims should be treated as reported allegations, not established findings.
Responses and Evidence to Watch
The next developments to watch are responses from OpenAI, DeepSeek and Moonshot, along with any fuller publication of The Information’s evidence. Technical details about access, the volume of model outputs and their role in training would help test the claim; none is provided in the public account described here.
OpenAI may also clarify whether it has taken action under its terms of service or changed controls on Astra access. Any such steps would need to be reported separately from the allegation itself. Until more evidence or official responses emerge, the status is developing, with no public resolution reported.
Key Questions
Which Chinese AI models are named in the report?
The report names Moonshot AI’s Kimi K3.1 Pro, released Oct. 23, and DeepSeek-V4.1-Flash-1017, released Oct. 17.
Have DeepSeek or Moonshot confirmed using Astra outputs?
No confirmation is included in the reporting summarized here. Reuters said DeepSeek, Moonshot and OpenAI did not immediately respond to its requests for comment.
Is AI distillation itself improper?
Not necessarily. Distillation is a common training technique. The disputed questions are whether outputs came from a competitor’s model, whether their use breached the provider’s terms, and how access was obtained.
What evidence has been made public?
The public account attributes the claim to people with direct knowledge, but does not give training figures, access records, methods or a technical analysis. The detailed Information report is paywalled, and the claim has not been publicly confirmed by the companies cited.
Source: ThorstenMeyerAI.com
