🔍 Read the full analysis: What To Know About China’s New AI Models And The GPT-6 Astra Report 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 while training recent models. Reuters said the three companies did not respond to its requests for comment. The report does not publicly provide evidence about the volume, method or source of the alleged outputs.
The Information reported October 28 that China’s DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra to train recent models, including Moonshot’s Kimi K3.1 Pro and DeepSeek’s V4.1 Flash update. The allegation could raise questions about model training and the terms governing access to AI services, but the companies have not publicly confirmed it, and the report’s public portion does not describe how the outputs were obtained.
Reuters summarized The Information’s report, which cited people said to have direct knowledge. It named Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, released October 17, as models that benefited from distilling Astra. The Information’s public opening described Astra as OpenAI’s most powerful released model and said Chinese AI labs were among a growing customer base seeking to extract its capabilities. The rest of its article is behind a paywall.
Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment. The source material also cites a Korean broadcaster’s October 29 summary, which attributed further details to unnamed sources and quoted an OpenAI spokesperson saying the company takes the issue seriously. Those details could not be independently checked against The Information’s full report. Neither model’s launch material, as described in the source, confirms the allegation: Moonshot credited Kimi’s gains to increased post-training compute and reinforcement learning, while DeepSeek’s update note did not discuss training data.
The report gives no public figures for the volume of Astra outputs, no detailed account of the prompts or collection process, and no evidence establishing whether fake or resold accounts were used. Those gaps matter: the claim of heavy reliance comes from the reporting, while the mechanism and scale remain unsubstantiated in the publicly available 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 Training Practices Matter
If the report is accurate, it would put a spotlight on the boundary between ordinary model distillation and using a competitor’s service to build a rival product. Distillation is a standard training technique: a smaller model learns from outputs produced by a larger one. The contested issues are who owns or controls the teacher model, how its outputs were obtained, and whether the use complied with the provider’s terms.
That distinction has practical consequences for AI developers, customers and model providers. Companies choosing a model may want to ask how its training data and outputs were sourced, while providers may tighten access controls or pursue contractual remedies. But the current report does not establish that a law was broken, nor does it provide enough public detail to judge the models’ training histories.
The source material lists prices of $10 per million input tokens and $50 per million output tokens for Astra, compared with $2 and $10 for Kimi K3.1 Pro, and $0.30 and $1.20 for DeepSeek’s model at peak hours. It also gives scores of 57, 46 and 32, respectively, on the same independent index. Those figures suggest lower listed prices alongside lower scores on that index; they do not show that the models are equivalent or explain the alleged training relationship.
Popular right now
Breville BES870XL Barista Express Espresso Machine, Brushed Stainless Steel
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
De'Longhi Magnifica Evo Automatic Espresso Machine with Milk Frother – Silver
As an affiliate, we earn on qualifying purchases.
Distillation and Earlier Allegations
In distillation, a teacher model generates answers to prompts, and a student model is trained to reproduce useful patterns in those answers. The technique can reduce the cost of serving a capable model and is commonly used by developers with their own systems. A U.S. government advisory cited in the source material describes distillation as a legitimate and useful AI research technique.
The dispute arises when outputs come from a rival’s model and their collection or use conflicts with service terms. OpenAI’s terms, according to the source material, prohibit using its outputs to build competing models. That could make the matter a contract question; alleged deception to obtain access could raise separate issues. The source author’s legal characterization is analysis, not a ruling or official finding.
The new report follows earlier public accusations. The source says that in September OpenAI connected a July campaign to extract hidden reasoning from its models to people associated with Moonshot AI, and that U.S. security agencies named both labs in an advisory. China’s commerce ministry called the accusations baseless. These are claims and responses from the parties, not independent confirmation of the current allegation. The source also references earlier Anthropic allegations involving DeepSeek, Moonshot and MiniMax; they provide context but do not prove the claims about Astra.
Evidence and Access Remain Unclear
The key allegation remains unconfirmed by the companies, and Reuters reported that OpenAI, DeepSeek and Moonshot did not immediately respond to its inquiries. The accessible reporting does not identify the accounts allegedly used, quantify the outputs, explain whether hidden reasoning was obtained, or show how much those outputs contributed to either model.
It is also unclear whether the alleged training, if it occurred as described, violated a contract or any law. OpenAI’s terms may be relevant, but deciding whether they applied and were breached would require facts not supplied in the public report. The models’ own launch descriptions do not independently settle the question.
Company Responses and Further Evidence
The next meaningful developments would be responses from OpenAI, DeepSeek or Moonshot, along with any additional reporting that documents the alleged training process or its scale. A public statement, technical evidence or legal filing could clarify whether the claim concerns ordinary access, prohibited use, or improperly obtained accounts. Until then, the reported relationship should be treated as an allegation rather than an established fact.
For organizations evaluating AI models, the immediate practical step is to ask vendors about training-data provenance and service terms, test models against their own tasks, and maintain a fallback option. The source material advises users to avoid unofficial API resellers, but does not report a change to the consumer apps or establish that users need to take action because of this allegation.
Key Questions
What does the report allege?
The Information alleges that DeepSeek and Moonshot relied heavily on GPT-6 Astra outputs to train recent models, including DeepSeek-V4.1-Flash-1017 and Kimi K3.1 Pro.
Have the companies confirmed the claim?
No. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment. The publicly available reporting does not include confirmation from the companies.
Is distillation itself improper?
Not by itself. Distillation is a common technique in which a student model learns from a teacher model’s outputs. Questions about this allegation concern the source of the outputs, how they were obtained, and whether their use complied with the provider’s terms.
Does the report establish that a law was broken?
No. The source material discusses possible contract and access issues, but the public report does not establish a legal violation. The available details are insufficient to determine whether any terms or laws were breached.
Source: ThorstenMeyerAI.com
