🔍 Read the full analysis: What Can We Infer From The Report On GPT-6 Astra And China’s New AI Models? 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 on October 28 that DeepSeek and Moonshot AI relied heavily on OpenAI’s GPT-6 Astra outputs to train recent models. Reuters said the companies did not immediately respond, and the public reporting does not specify the scale or method. The allegation matters because OpenAI’s terms prohibit using its outputs to build competing models, but the available evidence does not establish a breach.
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 recently released models, including DeepSeek-V4.1-Flash-1017 and Moonshot’s Kimi K3.1 Pro. Reuters repeated the report, but the companies had not immediately responded, and the public account does not provide evidence of the scale or method of the alleged training.
The Information’s report, by Jing Yang, Juro Osawa and Qianer Liu, cited people with direct knowledge, according to its public summary. The full article is paywalled. Reuters summarized its central claim as saying the two labs “relied heavily” on Astra outputs and that their latest models “benefitted from distilling” the OpenAI model. The claims remain attributed reporting, not a public finding supported by disclosed records.
The models named were DeepSeek-V4.1-Flash-1017, released October 17, and Kimi K3.1 Pro, released October 23. Neither company’s public release material, as described in the source material, confirms using Astra outputs. Moonshot’s launch post attributed Kimi’s gains to increased post-training compute and large-scale reinforcement learning; DeepSeek’s update note did not describe its training data.
Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to requests for comment. A separate summary by South Korea’s SBS quoted an OpenAI spokesperson as saying the company takes the issue very seriously and continues to work on it. That additional detail could not be checked against The Information’s full report, so it should be treated as secondary reporting.
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
The report puts attention on a distinction that matters to AI developers: distillation is a standard training technique, but using a competitor’s model outputs at scale may breach the provider’s terms. In distillation, a smaller “student” model learns from answers generated by a larger “teacher” model. Companies commonly use their own models in this way to create cheaper or more specialized versions.
OpenAI’s terms, according to the source material, prohibit using its outputs to build competing models. If the reported activity occurred and involved access obtained through deception or resale, it could raise contract and access-related questions. The report alone does not establish either how access was obtained or whether a legal violation occurred. Those issues depend on details that have not been made public.
The commercial stakes are also about cost and performance, not simply whether one model copied another. The source 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 at peak hours. It also reports scores of 57, 46 and 32 on the same independent index. Those figures suggest lower prices alongside lower scores on that index; they do not show that the models are equivalent or that distillation caused their results.
For companies selecting models, the allegation is a reason to ask vendors about training-data provenance, contractual assurances and fallback options. For users, the report does not establish an immediate change to the apps or services. The source advises avoiding unofficial, unusually cheap API resellers, but provides no evidence that a particular reseller was involved in this case.
Popular right now
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
De'Longhi Magnifica Evo Fully Automatic Espresso Machine with Manual Milk Frother
KitchenAid Artisan 5-Quart Tilt-Head Stand Mixer, Almond Cream
As an affiliate, we earn on qualifying purchases.
Earlier Disputes Over Model Outputs
The allegation follows earlier public disputes over attempts to extract capabilities from commercial models. In September, OpenAI said it linked a “core cluster” of a July campaign to obtain hidden reasoning from its models to people associated with Moonshot AI. The source also says U.S. security agencies named both DeepSeek and Moonshot in an advisory. Those are separate claims and actions; they do not independently verify that the labs used Astra outputs to train the specific October models.
The source material also describes an earlier Anthropic account, dated February 23, 2026, alleging that DeepSeek, Moonshot and MiniMax generated more than 16 million exchanges with Claude. The supplied text cuts off before explaining the allegation’s full scope, so that number should not be taken as evidence about the current Astra report. China’s commerce ministry called accusations against Chinese AI companies baseless, according to the same source material.
Distillation can involve ordinary model development or conduct that a provider alleges violates its rules, depending on whose outputs are used and how they are accessed. A U.S. government advisory cited in the source calls distillation “a legitimate and useful technique in AI research.” The central issue in this report is not the technique itself, but the claim that one company’s model outputs were used to build competing products.
Evidence Behind the Astra Allegation
The public record does not disclose how many Astra outputs were allegedly used, what prompts produced them, whether hidden reasoning was obtained, or how the outputs entered either model’s training pipeline. The Information cited unnamed sources, and its full report is behind a paywall. No public technical analysis or company documentation in the supplied material independently verifies the claim.
It is also unclear whether either lab used fake, resold or otherwise unauthorized accounts. That access method appears in the source’s general description of the dispute around large-scale distillation, not as a confirmed finding about these two models. The companies have not publicly confirmed or denied the specific allegation in the material provided. Neither model’s release notes, as summarized there, settle the question.
There is no public basis here to conclude that the alleged training violated a law, that the models’ performance came from Astra, or that users’ data or accounts are affected. The October report should be read as a serious but unverified allegation, distinct from prior claims about other campaigns and models.
Responses and Evidence to Watch
The next developments to watch are direct responses from OpenAI, DeepSeek and Moonshot, along with any technical evidence or further reporting that explains the alleged scale and method. A detailed account of account access, prompt records or training-data provenance would help determine whether the reported activity occurred as described.
For now, developers and buyers can ask model providers how they document training sources and how they handle third-party output restrictions. Companies evaluating the named models can test them on their own tasks and maintain a fallback, rather than treating a single index score or an unresolved allegation as decisive. The report does not specify a date for any investigation, company action or further disclosure.
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, including DeepSeek-V4.1-Flash-1017 and Kimi K3.1 Pro. The public summary cites people with direct knowledge; the claim has not been independently established in the material provided.
Have the companies confirmed the report?
No. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment. The supplied material contains no public confirmation or denial from the three companies about this specific allegation.
Is distillation itself prohibited?
Not inherently. Distillation is a common way to train a smaller model using outputs from a larger one. The reported concern is the alleged use of a competitor’s outputs in a way that may violate its terms, along with unanswered questions about how access was obtained.
Does the report prove the two models copied Astra?
No. It is an allegation based on unnamed sources in a report whose full text is paywalled. Public information cited here does not provide training records, a method, or a measurement of how much Astra output was used.
Does this change anything for users right now?
The supplied reporting does not identify an immediate change to the apps or services. It also does not establish an impact on user accounts or data. Business users may wish to ask vendors about training-data provenance while the claims remain unresolved.
Source: ThorstenMeyerAI.com
