🔍 Read the full analysis: Did China’s Newest AI Models Learn From GPT-6 Astra? The Report, In Brief on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
The Information reported that DeepSeek and Moonshot AI relied heavily on OpenAI’s GPT-6 Astra outputs to train recent models, citing people with direct knowledge. Reuters reported that the three companies did not immediately respond; the public reporting offers no figures or details showing 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 models released this month. Reuters summarized the report within hours, but the claim has not been publicly confirmed: Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment.
The models named are Moonshot’s Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, released October 17. The Information cited people with direct knowledge, according to Reuters’ account. The original article’s public opening describes Astra as a strong draw for Chinese AI labs seeking to extract its capabilities; the remainder is behind a paywall, limiting what can be independently checked from the public record.
The report does not provide the number of Astra outputs involved, a detailed account of the training method, or evidence showing how any accounts used to access Astra were obtained. Moonshot’s launch post attributed Kimi K3.1 Pro’s gains to larger post-training compute and large-scale reinforcement learning. DeepSeek’s update note, as described in the source material, does not discuss training data. Those company statements do not confirm or disprove the reported use of Astra outputs.
Reuters said all three companies had not immediately responded. Separately, South Korean broadcaster SBS said an OpenAI spokesperson described the issue as one the company takes seriously and said it continues working on it. That wording is relayed through SBS, rather than independently confirmed here against the original report. No company has publicly confirmed the specific training allegation in the material 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 Provenance Matters
The allegation concerns more than whether one model can imitate another. Training on a competitor’s outputs at scale could reduce the cost and time needed to reproduce some capabilities, while raising questions about access controls, contract terms and the provenance of commercial AI systems. The report does not establish that this occurred unlawfully, or say whether the outputs materially changed either model’s performance.
For businesses choosing models, the immediate practical issue is uncertainty about training provenance, not a confirmed change to the products. Buyers can ask providers what information they disclose about training, evaluate models on their own tasks and retain a fallback where a service is business-critical. The source article also advises users to avoid unofficial API resellers; that is general caution, not evidence that any particular reseller was involved in this case.
The reported allegation also matters to AI providers because their terms and technical protections may be tested when customers use model outputs to train competing systems. Yet the legal and policy implications depend on facts not made public, including the scale of any use, how access was obtained, and the terms that applied. The current report leaves those questions open.
Products Worth Considering
Apple AirPods Pro 3 Wireless Earbuds
DJI Osmo Pocket 3 Vlogging Camera with 1-Inch CMOS Sensor and 4K Video
DJI Mic 3 Wireless Microphone System (2 Transmitters, 1 Receiver, Charging Case)
As an affiliate, we earn on qualifying purchases.
Distillation and Earlier Allegations
Distillation is a standard model-training technique: a smaller model is trained using responses generated by a larger one, often called the teacher. AI labs use the approach to make systems that are cheaper or faster to run. The U.S. government advisory cited in the source material calls distillation a legitimate and useful research technique. The controversy is not the technique by itself, but the possibility that a lab used a rival’s model at large scale in ways that breached service terms or relied on deceptive access. Those details have not been established in the current reporting.
The report arrives amid prior public disputes over model extraction. The source material says OpenAI linked a July campaign to obtain hidden reasoning from its models to people associated with Moonshot AI, and that U.S. security agencies named DeepSeek and Moonshot in an advisory. These are separate claims and allegations; they do not independently prove that either lab used Astra outputs to train the October models. China’s commerce ministry has called accusations of this kind baseless, according to the source material.
The source article compares listed prices and scores on an independent index: Astra at $10 per million input tokens and $50 per million output tokens, Kimi K3.1 Pro at $2 and $10, and DeepSeek’s model at $0.30 and $1.20 at peak hours. Their reported index scores are 57, 46 and 32, respectively. These figures are presented as the article’s comparison, not proof of training influence; the models are cheaper but do not score equally on that index. The source does not specify further details about the index or its test set.
Evidence Behind the Astra Claim
The central allegation remains unverified publicly. The available account relies on The Information’s unnamed sources and Reuters’ summary; the underlying details are largely behind a paywall. There are no disclosed figures for the volume of outputs, no described method for collecting them, and no public evidence identifying whether access came through ordinary accounts, resellers or another route.
It is also unclear whether Astra outputs were a major component of either model’s training, whether any use breached OpenAI’s terms, and whether the practice affected model performance. The companies’ silence in Reuters’ initial reporting is not confirmation. Nor do the launch notes from Moonshot and DeepSeek settle the question: they provide some account of development but do not address the allegation directly.
Company Responses and Evidence
The next developments to watch are direct responses from OpenAI, DeepSeek and Moonshot, along with any additional evidence from The Information or other reporting. A clearer account would need to establish what outputs were used, at what scale, how they were accessed and what terms governed that access. Until then, the report should be treated as an attributed allegation, not a confirmed account of the models’ training.
For users and organizations, there is no confirmed app or product change described in the reporting. Companies evaluating these systems can monitor provider statements and assess model quality and reliability on their own workloads. Any future legal, policy or product response will depend on facts that have not yet been made public.
Key Questions
Which Chinese AI models are named in the report?
The Information report, as summarized by Reuters, names Moonshot AI’s Kimi K3.1 Pro and DeepSeek-V4.1-Flash-1017. The models were released on October 23 and October 17, respectively, according to the source material.
Has OpenAI confirmed that the models were trained on Astra outputs?
No. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment. The report’s claim has not been publicly confirmed by the companies in the source material.
What does distillation mean?
Distillation is a training method in which a smaller model learns from responses generated by a larger model. It is a common technique; the dispute in this report concerns the alleged use of a competitor’s outputs and the scale and means of access, which remain unclear.
Does the report prove that OpenAI’s terms were violated?
No. The source material says OpenAI’s terms forbid using its outputs to build competing models, but the report does not establish what access or training occurred or whether the terms were breached. The legal question depends on facts not yet public.
Source: ThorstenMeyerAI.com
