🔍 Read the full analysis: Did GPT-6 Astra Influence China’s Newest AI Models? The Evidence In The Report 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 GPT-6 Astra outputs to train recent models; Reuters summarized the allegation, citing the publication. The public evidence is limited: the report relies on unnamed sources, and the companies have not confirmed the claim or described how any outputs were obtained.
The Information reported October 28 that Chinese AI companies DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra to train recent models, a claim that could intensify disputes over how AI labs use competitors’ systems. Reuters summarized the report, but none of the three companies had confirmed the allegation in Reuters’ account, which cited unnamed sources and did not detail the evidence publicly.
The models identified in the coverage are Moonshot’s Kimi K3.1 Pro and DeepSeek-V4.1-Flash-1017. The Information said both benefited from distilling Astra, according to Reuters’ summary. Kimi K3.1 Pro was released October 23, while DeepSeek’s V4.1 Flash update was launched October 17, according to the source material.
The public record does not establish how many Astra outputs were used, how the labs obtained them, or what share of either model’s training came from those outputs. The Information’s article is largely behind a paywall, and its publicly available opening describes Chinese AI labs as a growing customer base seeking to extract Astra’s capabilities. Reuters reported that OpenAI, DeepSeek and Moonshot did not immediately respond to its request for comment.
Moonshot’s launch post attributed Kimi K3.1 Pro’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. Those statements do not independently confirm or refute the reported use of Astra outputs.
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 matters because distillation is a routine training method, but using another company’s model at scale may raise contractual and access questions. OpenAI’s terms, according to the source material, prohibit using its outputs to build competing models. Whether those terms were breached would depend on facts not established in the public reporting, including how access was obtained and what training use occurred. Any legal conclusions remain unconfirmed.
The reported models are also positioned at different price and performance levels. The source material 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 during peak hours. It gives scores of 57, 46 and 32 on the same independent index, respectively, without identifying the index in the supplied text. These figures suggest a price-performance trade-off, not that the less expensive models match Astra.
For businesses choosing AI systems, the story is a reminder that model provenance and terms of use can form part of vendor due diligence, alongside cost, performance and reliability. For users, the report does not establish that the named apps have changed or that their outputs are affected. The source advises caution with unofficial, low-cost API resellers; it does not document a specific reseller’s role in this case.
Products Worth Considering
Apple MacBook Pro 16-inch (2021) – M1 Pro Chip, 10-Core CPU, 16-Core GPU, 16GB RAM, 1TB SSD, Space Gray
Apple MacBook Pro 16.2-inch with M5 Pro, 18-core CPU, 20-core GPU, 48GB Memory, 1TB SSD, Wi-Fi 7, Space Black
LG UltraGear 45GX900A-B 45-Inch Curved OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Earlier Allegations and Distillation Rules
Distillation trains a model using answers generated by another model, often called a teacher. Labs commonly use the method to make smaller or less costly versions of their own systems. A U.S. government advisory described distillation as a legitimate and useful AI research technique, according to the source material. The dispute in this case is not the technique alone, but the reported use of a rival’s outputs and the means by which those outputs may have been accessed.
The latest report follows earlier public accusations. In September, OpenAI linked what it called a “core cluster” of a July campaign to attempts to extract hidden reasoning from its models, and associated individuals with Moonshot AI, according to the supplied account. U.S. security agencies also named DeepSeek and Moonshot in an advisory. China’s commerce ministry called the accusations baseless, the source material says. Those prior claims provide context but do not prove that either company used Astra outputs to train the models now named.
The source also describes a February 2026 Anthropic allegation involving DeepSeek, Moonshot and MiniMax, but the supplied passage ends before giving its full details. Separately, Korean broadcaster SBS reported that the labs had shifted from targeting Anthropic’s Claude to OpenAI after Anthropic tightened defenses, and quoted an OpenAI spokesperson as saying the company took the issue seriously. The source’s author said those details could not be checked against The Information’s original reporting, so they should be treated as secondary, unverified details, not established facts.
“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
Key Evidence Still Missing
The allegation has not been independently established in the public material supplied. The Information cited people with direct knowledge, but the sources are unnamed, the full article is paywalled, and no underlying records are presented here. Reuters’ report says the companies did not immediately respond; that is not the same as a denial or confirmation.
There are also no public figures for the volume of Astra outputs, the prompts used, the account holders involved, or the role those outputs played in either model’s training. It is unclear whether access involved ordinary customer accounts, resold access, deceptive accounts or another method. The source says OpenAI’s terms bar using its outputs to build competing models, but whether any particular term was violated cannot be determined from the available reporting.
The companies’ published training explanations offer only partial information: Moonshot cited compute and reinforcement learning, while DeepSeek’s update note did not address training data. Neither public statement, as described here, settles the specific allegation.
Responses and Evidence to Watch
The next key development would be a substantive response from OpenAI, DeepSeek or Moonshot, or the release of further reporting that supports or challenges the claim with verifiable details. Public information about account access, model-output volumes or training methods could clarify whether the reported activity occurred and how extensive it was. No formal investigation or enforcement action is established in the supplied material.
Companies evaluating these models can continue checking vendor documentation, contract terms and performance on their own tasks; the source recommends maintaining a fallback system. Those are procurement considerations, not conclusions that the reported models are unreliable. For users, the practical status is unchanged on the available evidence: the report alone does not show that the apps or services have altered their behavior.
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, including DeepSeek-V4.1-Flash-1017 and Moonshot’s Kimi K3.1 Pro. The claim is based on unnamed sources in the coverage summarized by Reuters.
Have DeepSeek, Moonshot or OpenAI confirmed the claim?
No confirmation is reported in the supplied material. Reuters said the three companies did not immediately respond to its request for comment; that does not establish whether they later responded or what their positions are.
Is distillation itself improper?
Not inherently. Distillation is a common technique for training smaller models on a larger model’s outputs. The disputed issues here would include whether a rival’s outputs were used in breach of terms and whether access was obtained through improper means. The available reporting does not resolve those questions.
Does the report prove the named models copied GPT-6 Astra?
No. The report alleges heavy reliance on Astra outputs, but the public material supplied contains no training records, output counts or detailed method. The claim remains unconfirmed by the companies and is not independently proven by the information provided.
What should model users take from the story?
The report does not establish a change to the named products or show that their apps are affected. Organizations may review vendor terms, provenance disclosures and task-specific performance; users should distinguish the reported allegation from confirmed findings.
Source: ThorstenMeyerAI.com
