🔍 Read the full analysis: China’s New AI Models: What The GPT-6 Astra Distillation Report Shows 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 October 28 that DeepSeek and Moonshot AI relied heavily on OpenAI’s GPT-6 Astra outputs to train recent models. The claim, based on unnamed sources, has not been confirmed by the companies; public details about the scale and method are limited.
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, a claim that could raise questions about model providers’ terms and how training data is sourced. The report cites people with direct knowledge, but none of the three companies has publicly confirmed it, and the available accounts give few details about the alleged activity.
Reuters summarized the report as saying the models that benefited were Moonshot’s Kimi K3.1 Pro and DeepSeek-V4.1-Flash-1017. Kimi K3.1 Pro was released October 23, according to the source material; DeepSeek’s update was launched October 17. The Information’s public article opening described Astra as a powerful model with strong demand since its release in early September, but its fuller account is paywalled.
Reuters reported that OpenAI, DeepSeek and Moonshot did not immediately respond to requests for comment. The public reporting does not establish how many Astra outputs were involved, how they were obtained, or how much they contributed to either model. Moonshot’s launch post attributed Kimi’s gains to substantially larger post-training compute and large-scale reinforcement learning. DeepSeek’s update note, as described in the source material, did not discuss training data.
A separate account from South Korea’s SBS, published October 29, said the sources described a shift from using Anthropic’s Claude to OpenAI models after Anthropic tightened defenses. SBS also quoted an OpenAI spokesperson saying the company takes the issue seriously and continues working on it. Those details have not been independently checked against The Information’s full report and should be treated as additional attributed reporting, not as established findings.
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
If the report is accurate, it would put a spotlight on the boundary between ordinary model distillation and using a competitor’s service to train a rival system in ways its terms prohibit. Distillation is a common technique: a smaller model learns from examples produced by a larger one. The disputed issue is not the technique by itself, but whether Astra outputs were gathered at scale and used in a manner that breached OpenAI’s conditions or involved deceptive access.
The commercial stakes are visible in the prices and benchmark figures cited by the source material. Astra is listed 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 during peak hours. On the same independent index, the three are reported to score 57, 46 and 32, respectively. These figures indicate different price and score points, not equivalent performance, and the available material does not identify the index or its measurement details.
For businesses selecting models, the story is a reason to ask vendors about training-data provenance, terms compliance and testing, while keeping alternatives available. For users, the report does not establish that the apps or services have changed. The source advises caution with unofficial, low-cost API resellers, but provides no evidence that a particular reseller was involved.
Popular right now
De'Longhi Magnifica Evo Automatic Espresso Machine with Milk Frother – Silver
De'Longhi Magnifica Evo Fully Automatic Espresso Machine with Manual Milk Frother
Breville BES870XL Barista Express Espresso Machine, Brushed Stainless Steel
As an affiliate, we earn on qualifying purchases.
Earlier Disputes Over Model Extraction
The allegation follows earlier public disputes about attempts to extract capabilities from commercial AI systems. In September, OpenAI said it had linked a “core cluster” in a July campaign to obtain hidden reasoning from its models to individuals associated with Moonshot AI. The source material also says U.S. security agencies named DeepSeek and Moonshot in an advisory. These are claims and actions attributed to OpenAI and government agencies; they do not independently prove the latest report’s account. China’s commerce ministry has called accusations against Chinese AI firms baseless, according to the source.
The government advisory described distillation as a legitimate and useful technique in AI research. Labs commonly train smaller models on outputs from their own larger models. A provider may object when a rival uses its outputs contrary to service terms, particularly if access was obtained through false or resold accounts. Whether any such conduct occurred here is not established by the public reporting. The distinction matters: a dispute over contractual terms or account access is not automatically proof of a broader legal violation.
The value of model outputs can also depend on whether they include detailed reasoning traces. The source material cites an Anthropic experiment in which a model trained on full reasoning transcripts from Claude Opus 4.8 scored 80.4% on SWE-bench Verified, compared with 71.8% when trained on summarized transcripts. That is Anthropic’s own reported experiment, not evidence about Astra or the Chinese models in this report.
“DeepSeek and Moonshot “relied heavily on outputs” from GPT-6 Astra to train their latest models.”
— The Information, as summarized in the report
What the Public Evidence Leaves Open
The central claim remains unconfirmed by OpenAI, DeepSeek and Moonshot. The available material does not disclose the number of prompts or outputs involved, the accounts or access methods allegedly used, the dates of any activity, or how much the alleged training influenced the released models. The Information’s report relies on unnamed sources, and the article’s fuller detail is behind a paywall.
It is also unclear whether the alleged activity, if it took place, violated specific contractual terms or involved fraud. The source material notes that OpenAI’s terms prohibit using its outputs to build competing models, but determining whether those terms applied and were breached would require facts not publicly provided. No regulator or court finding on this allegation is cited.
The companies’ public explanations do not resolve the issue: Moonshot attributed its model’s gains to compute and reinforcement learning, while DeepSeek’s update note reportedly said nothing about its training data. Those statements neither confirm nor refute the reported use of Astra outputs.
Responses and Evidence to Watch
The next useful development would be direct responses from OpenAI, DeepSeek and Moonshot, along with any additional reporting that clarifies the sources, scale and method behind the allegation. The public account may also become more detailed if The Information’s full report or further documentation is made available.
Companies evaluating AI services can review providers’ statements about data provenance and contractual compliance, test models on their own tasks, and maintain a fallback provider. Those steps are practical procurement safeguards, not proof that a specific model was trained improperly. For users, there is no confirmed change to the named products based on the reporting so far; the allegation and any investigation or response remain developing.
Key Questions
What did the report allege?
The Information reported that DeepSeek and Moonshot AI relied heavily on GPT-6 Astra outputs to train recent models. Reuters identified Kimi K3.1 Pro and DeepSeek-V4.1-Flash-1017 as the models involved, according to the report.
Have the companies confirmed the allegation?
No. Reuters reported that OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment. The public information cited here does not include confirmation from any of them.
Is distillation itself improper?
No. Distillation is a common model-training technique, and the government advisory cited in the source material calls it legitimate and useful. The concern in this report is whether a rival’s outputs were used at scale in a way that violated service terms or involved improper access.
What evidence has been made public?
The report, as summarized by Reuters, cites people with direct knowledge. Public accounts do not provide the alleged prompt volume, access method, training contribution or documentary evidence. The companies have not confirmed the claim.
Does this change anything for users now?
The reporting does not establish a change to the named apps or models. It raises questions about training-data provenance and service terms, but any product or policy consequences remain unknown.
Source: ThorstenMeyerAI.com
