🔍 Read the full analysis: What Does The Distillation Report Say About China’s AI Models? on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
The Information reported that DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra to train recent models. The allegation is based on unnamed sources, and the public report does not provide figures or explain how the outputs were obtained; the three companies had not confirmed or denied it to Reuters.
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, including Kimi K3.1 Pro and DeepSeek V4.1 Flash. The allegation, attributed to people with direct knowledge, has not been publicly substantiated with data or a detailed account of the training; Reuters said the three companies did not immediately respond to requests for comment.
Reuters summarized The Information’s report as saying the labs’ latest models “benefitted from distilling Astra.” The models named were Moonshot’s Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, launched October 17. The Information’s public article opening described Chinese AI labs as a growing customer base for Astra; the remainder of its report is paywalled, limiting what can be checked from the public record.
The account does not disclose how many Astra outputs were used, what prompts or training methods were involved, or how the labs obtained access. Reuters reported that OpenAI, DeepSeek and Moonshot did not immediately respond. The models’ own release materials offer other explanations for their development: Moonshot credited Kimi K3.1 Pro’s gains to “substantially larger post-training compute” and large-scale reinforcement learning, while DeepSeek’s update note did not describe its training data. Those statements do not independently confirm or disprove the reported use of Astra outputs.
A Korean broadcaster, SBS, summarized further details on October 29, including a claim that Chinese labs had previously focused on Anthropic’s Claude and shifted toward OpenAI after Anthropic tightened defenses. SBS also quoted an OpenAI spokesperson saying the company takes the issue very seriously and continues to work on it. Those details could not be checked against the paywalled original from the material available, and the companies had not publicly confirmed the central allegation.
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 Developers
If the report is accurate, the dispute concerns both competitive advantage and the terms under which a model provider’s service may be used. Distillation is a standard way to train a smaller model to imitate a larger one’s answers. The U.S. government advisory cited in the source describes it as “a legitimate and useful technique in AI research.” The contested issue is whether a rival’s outputs were gathered and used at scale in ways that breach its rules, not whether distillation is inherently improper.
OpenAI’s terms prohibit using its outputs to build competing models, according to the source material. That makes the allegation potentially a contract dispute; claims that access was obtained through fake or resold accounts could raise separate questions about how the service was used. No public evidence in the report establishes that such access methods were used by either lab, or determines whether any law was broken.
The commercial comparison helps explain the attention around the claim, but does not establish model equivalence. The source lists 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. On the same independent index cited by the source, the models score 57, 46 and 32, respectively. These figures suggest lower listed costs alongside lower scores on that index; they do not establish how the models perform on every task or whether distillation caused their results.
For organizations choosing AI systems, the report is a reminder that model provenance and provider terms can matter alongside price and benchmark scores. The source recommends checking model provenance, maintaining a fallback, and testing systems on an organization’s own tasks. Those are risk-management considerations, not a finding that the named models are unusable or improperly trained.
Products Worth Considering
Amazon Kindle Scribe Colorsoft 64GB (1st Generation)
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
Apple iPad Air (4th Generation, 10.9-inch, Wi-Fi, 256GB) – Space Gray
As an affiliate, we earn on qualifying purchases.
Earlier Claims About Model Extraction
The October report follows earlier public allegations about attempts to extract capabilities from commercial models. In September, OpenAI said it had linked a “core cluster” of a July campaign aimed at obtaining hidden reasoning from its models to individuals associated with Moonshot AI. The source also says U.S. security agencies named DeepSeek and Moonshot in an advisory. These are claims and assessments reported by one side, not independent confirmation of the current allegation that Astra outputs were used to train the two named models.
The source describes a separate February 23, 2026, Anthropic account involving DeepSeek, Moonshot and MiniMax and exchanges with Claude. The excerpt provided here cuts off before completing that account, so it does not support further detail. China’s commerce ministry has called accusations against the labs baseless, according to the source. This background shows a continuing dispute over access to AI systems and model extraction, but it does not resolve what happened in the Astra case.
In distillation, a “teacher” model generates answers to prompts and a “student” model is trained to imitate them. Providers may restrict the use of their outputs in competing services. The source also 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 experiment, not evidence about DeepSeek or Moonshot’s training methods.
“DeepSeek and Moonshot “relied heavily on outputs from OpenAI’s powerful GPT-6 Astra model to train their latest models.””
— The Information, as summarized by Reuters
What the Public Record Lacks
The central claim remains unconfirmed by the companies in the accounts provided. The Information relied on unnamed sources, and the public portion of its story does not supply the evidence behind the claim. There are no disclosed totals for Astra queries or outputs, no description of the accounts used, and no technical analysis showing how much of either model’s training depended on Astra.
It is also unclear whether any use violated OpenAI’s terms, whether access was obtained through deceptive or resold accounts, and what remedies or consequences might follow. Reuters said the companies did not immediately respond; that is not a confirmation, denial or final position. The reported benchmark and price figures do not independently verify a model’s training history. No legal finding or public investigation result is identified in the supplied material.
Company Responses and Evidence
The next developments to watch are responses from OpenAI, DeepSeek and Moonshot, along with any further reporting that supplies evidence beyond the initial anonymous-source account. More detail about query volumes, access channels or training methods would help readers judge the scale and nature of the alleged use. Public documentation from the labs could also clarify how they describe the models’ training, although existing launch notes do not settle the allegation.
For now, the reported dispute has not changed the models’ availability, according to the source material. Organizations using them can review service terms and provenance information and test performance on their own workloads. The report provides no basis for predicting regulatory action, legal outcomes or future model releases; those remain open questions.
Key Questions
Which models are named in the report?
The report names Moonshot AI’s Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, launched October 17. The Information said the models benefited from distilling OpenAI’s GPT-6 Astra; neither company had publicly confirmed that claim in the source material.
What does distillation mean in this dispute?
Distillation trains a smaller “student” model to imitate answers produced by a larger “teacher” model. The technique is used in AI research and development. The dispute is about the reported use of a rival provider’s outputs and whether that use complied with its service terms.
Has OpenAI confirmed the allegation?
No confirmation is described. Reuters reported that OpenAI, DeepSeek and Moonshot did not immediately respond to requests for comment. SBS separately reported a general statement from an OpenAI spokesperson, but the source material says that detail could not be checked against the paywalled original.
Does the report prove that the labs broke the law?
No. The public account does not establish how the outputs were obtained or whether any law was broken. OpenAI’s terms reportedly forbid using its outputs to build competing models, which could make the issue a contract question; alleged access methods and legal consequences remain unverified.
What should AI users take from the report?
The report is an allegation, not a finding about the quality or legality of either model. Organizations may want to check model provenance and terms and evaluate systems on their own tasks. The source says the models remain available; it reports no immediate change for users.
Source: ThorstenMeyerAI.com
