🔍 Read the full analysis: What The Distillation Findings Say About GPT-6 Astra And China’s AI Models on ThorstenMeyerAI.com
Get the latest gadgets 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 on October 28 that DeepSeek and Moonshot AI relied heavily on OpenAI’s GPT-6 Astra outputs to train recent models. Reuters said the three companies did not respond to its requests for comment; the public evidence does not specify the scale or method. Distillation is a common training technique, but using a rival’s outputs may breach its terms, depending on what happened.
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. Reuters summarized the allegation within hours, but reported that OpenAI, DeepSeek and Moonshot did not respond to its requests for comment, leaving the claim unconfirmed by the companies.
The models named in the report are Moonshot’s Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, released October 17. Reuters’ account of The Information’s reporting says both models “benefitted from distilling Astra.” The Information’s public article opening describes Chinese AI labs as a growing customer base for Astra; the rest of its report is paywalled, limiting what can be independently checked from the public record.
The reporting does not provide the number of Astra outputs involved, how those outputs were obtained, or evidence showing how much they affected either model’s capabilities. The Information cited people with direct knowledge, according to Reuters, but the sources are not named in the public accounts. The model makers’ own public descriptions do not settle the matter: Moonshot’s launch post attributed Kimi’s gains to more post-training compute and large-scale reinforcement learning, while DeepSeek’s update note did not describe its training data.
A separate account by South Korea’s SBS on October 29 added that the labs had previously focused on Anthropic’s Claude and turned to OpenAI after Anthropic tightened defenses. SBS also attributed a statement to an OpenAI spokesperson saying the company takes the issue seriously and continues to work on it. Those details have not been independently verified against The Information’s paywalled report in the supplied material; they should be treated as SBS’s account, not as independently established facts.
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 could sharpen a dispute over who may use a model’s outputs and under what conditions. Distillation—training a model to imitate responses from a more capable model—is widely used, including by labs to make smaller versions of their own systems. OpenAI’s terms prohibit using its outputs to build competing models, according to the source material. Whether the alleged conduct violated those terms would depend on what was done, how access was obtained and which terms applied.
The issue has practical consequences for companies choosing AI systems. A model may be cheaper than a leading rival without matching its performance, and its training provenance may matter to a buyer’s legal, security and procurement reviews. The cited source lists Astra at $10 per million input tokens and $50 per million output tokens, 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 one independent index. Those figures are the source’s comparison; the index and test details are not provided here, so they do not establish performance on every task.
For users, the report does not by itself show that the named apps are unsafe or that their current behavior will change. For businesses, it is a reason to ask vendors about training-data provenance, keep alternatives available and test models on their own workloads rather than relying on broad rankings.
Products Worth Considering
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
Samsung 27-Inch Odyssey QD-OLED G8 4K 240Hz Gaming Monitor
Apple AirPods Pro 3 Wireless Earbuds
As an affiliate, we earn on qualifying purchases.
Distillation and Earlier Allegations
Distillation is a standard machine-learning technique. A teacher model generates answers to prompts, and a student model is trained to reproduce them. The U.S. advisory referenced in the source material describes distillation as a legitimate and useful technique in AI research. The controversy is not the technique alone: it concerns allegations of extracting a competitor’s capabilities at scale, potentially through access that violates service terms.
The October report follows earlier public accusations, though those do not prove the new claim. In September, OpenAI said it had linked a “core cluster” from a July campaign targeting its models’ hidden reasoning to individuals associated with Moonshot AI. A U.S. advisory also named DeepSeek and Moonshot, according to the source material. China’s commerce ministry called the accusations baseless. These are statements and allegations by the parties involved, not independent findings that establish what happened in the training of the two October models.
The distinction matters because the value of distillation may depend on what the teacher provides. The source describes 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 Astra or the Chinese models. It illustrates why detailed outputs may be valuable, while leaving the central allegations in this report unresolved.
“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
Evidence Behind the Astra Claim
The public accounts do not establish how many Astra outputs were used, how the labs obtained access, or whether any accounts were fake, resold or otherwise improperly used. The allegation rests on unnamed sources cited by The Information, and the underlying report is largely paywalled. Reuters said the three companies did not respond to its requests for comment; that is not the same as a denial or confirmation.
It is also unclear whether the alleged outputs were a major cause of the models’ performance, whether the use breached applicable terms, or whether any legal action or enforcement process is underway. The source’s pricing and index-score comparisons do not answer those questions. The separate SBS account has not been independently checked against the original report in the supplied material.
Responses and Provenance Checks
The immediate next development would be a response from OpenAI, DeepSeek or Moonshot that addresses the reported use of Astra outputs, or further reporting that provides evidence about the method and scale. No public investigation outcome, legal filing or enforcement action is identified in the source material. Any later account should distinguish a company’s allegation from independently verified evidence.
Companies evaluating the models can meanwhile ask providers about training-data provenance and account controls, review contractual terms, test performance on their own tasks and maintain fallback systems. The source also advises users to avoid cheap, unofficial API resellers. Those are precautions, not findings that the named models or official apps are currently compromised.
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, released October 17.
Have OpenAI, DeepSeek or Moonshot confirmed the allegation?
Not in the cited public reporting. Reuters said the three companies did not respond to its requests for comment. That does not establish whether they later responded or what their position is now.
Is distillation itself improper?
No. Distillation is a standard training technique and can be used legitimately, including to create smaller versions of a lab’s own models. The dispute concerns the alleged use of a competitor’s outputs and whether the way they were obtained or used breached applicable terms.
Does the report prove the Chinese models copied Astra?
No. It reports an allegation based on unnamed sources. Public accounts do not provide the underlying data, the scale of any use or a method for independently verifying its effect on the models.
What should AI buyers do while the claim is unresolved?
Buyers can ask providers about training-data provenance, check contractual and security terms, test systems on their own tasks and keep a fallback option. The report alone does not establish that the models are unsafe or unsuitable.
Source: ThorstenMeyerAI.com
