🔍 Read the full analysis: Did China’s Newest AI Models Learn From GPT-6 Astra? What The Distillation Report Shows, And What It Doesn’t 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 reported that none of the three companies immediately responded, and the public evidence does not describe the scale or method of the alleged training. Distillation is a common technique; the dispute concerns use of a rival’s outputs and whether access violated OpenAI’s terms.
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 DeepSeek-V4.1-Flash-1017 and Moonshot’s Kimi K3.1 Pro. Reuters relayed the report, but the public account rests on unnamed sources, and none of the three companies had confirmed or denied the claim in Reuters’ initial coverage.
The Information’s report, published at 6 p.m. Pacific time, said Astra had become a growing source of training outputs for Chinese AI labs seeking to reproduce some of its capabilities. Reuters summarized the central allegation as the labs having “relied heavily” on Astra outputs. The Information named Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, released October 17, as models that benefited from distillation.
The public reporting does not provide a number of prompts or outputs, explain how the labs allegedly accessed Astra, or specify how much any such data contributed to either model. The original article is largely behind a paywall. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to requests for comment. Moonshot’s release post attributed Kimi K3.1 Pro’s gains to substantially more post-training compute and large-scale reinforcement learning; DeepSeek’s update note did not describe its training data.
South Korea’s broadcaster SBS, in coverage published October 29, added that sources said Chinese labs had previously focused on Anthropic’s Claude and shifted toward OpenAI after Anthropic tightened defenses. SBS also quoted an OpenAI spokesperson as saying the company takes the issue seriously and continues working on it. Those details could not be independently checked against the full Information report in the source account, so they remain attributed reporting, not 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 Sources Matter
If the allegation is substantiated, it would sharpen a growing dispute over who may use a model’s outputs to train a competing system. Distillation can reduce the cost of building a capable model by using a stronger system’s answers as training material. That can benefit developers and customers, but it may also let a competitor capture some value from a model built with another company’s investment.
The commercial gap described in the source is substantial, though it does not establish that the models are equivalent. At the cited prices, Astra costs $10 per million input tokens and $50 per million output tokens; Kimi K3.1 Pro is listed at $2 and $10, while DeepSeek’s model is $0.30 and $1.20 at peak hours. On the same independent index cited by the source, the models score 57 for Astra, 46 for Kimi and 32 for DeepSeek. Those figures suggest lower prices alongside lower scores on that index; they do not establish why the models perform as they do.
For companies choosing AI systems, the story is a reminder to ask vendors about training-data provenance and contractual assurances, and to test models on their own workloads. The source also advises keeping a fallback model and avoiding cheap, unofficial API resellers. It reports no immediate change to consumer-facing apps.
Products Worth Considering
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Black
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
As an affiliate, we earn on qualifying purchases.
Distillation and Earlier Allegations
Distillation is a standard training approach: a “student” model learns from answers generated by a “teacher” model. Labs commonly use their own larger models to create smaller or less expensive versions. The source quotes a U.S. government advisory describing distillation as “a legitimate and useful technique in AI research.” The method itself is not evidence of wrongdoing.
The concern in this report is the alleged use of a competitor’s model at scale, potentially through access that violates service terms. OpenAI’s terms, according to the source, prohibit using its outputs to build competing models. Whether that provision applies to the reported conduct, and whether any access involved deception or resold accounts, cannot be determined from the available reporting. Any legal conclusion would depend on facts and applicable law.
The report follows earlier public allegations about attempts to extract model capabilities. The source says OpenAI linked a July campaign to extract hidden reasoning to people associated with Moonshot AI in a September account, while U.S. security agencies named both DeepSeek and Moonshot in an advisory. China’s commerce ministry called those accusations baseless. These are prior claims and responses; they do not independently prove the new allegation about Astra or either October model.
“DeepSeek and Moonshot “relied heavily on outputs” from GPT-6 Astra to train their latest models.”
— The Information, as summarized by Reuters
Evidence Behind the Astra Claim
The main unresolved issue is the evidence supporting the reported training link. The public material identifies unnamed sources said to have direct knowledge, but it does not disclose the number of Astra outputs involved, the prompts used, the accounts involved, or how the information was verified. The companies’ lack of an immediate response is not confirmation.
It is also unclear whether any Astra outputs were incorporated into the released models, how much they may have influenced performance, or whether the reported conduct breached a contract or law. The companies may offer different explanations or supporting information. Until more evidence or responses become public, the claim should be treated as reported but unconfirmed.
Company Responses and Evidence
The next developments to watch are responses from OpenAI, DeepSeek and Moonshot, along with any fuller account of the methods and evidence behind The Information’s reporting. Further details could clarify whether the allegation concerns direct training on model outputs, the scale of any activity, or access practices that may have violated service terms.
For now, the release notes and public reporting do not resolve those questions. Businesses can review their model-vendor documentation and test alternatives on their own tasks; users can avoid unofficial API resellers. Any assessment of responsibility or competitive impact should wait for more specific evidence and responses from the companies.
Key Questions
Did DeepSeek and Moonshot confirm that they trained on GPT-6 Astra outputs?
No public confirmation is included in the reporting. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment.
Which models does the report name?
The report names Moonshot’s Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, released October 17.
Is model distillation itself improper?
No. Distillation is a common training technique and can be used to make smaller models from a lab’s own larger systems. The allegation concerns using a rival’s outputs and the circumstances under which they may have been obtained.
Does the report show that the Chinese models match Astra?
No. The source cites different scores for the three models on one independent index: 57 for Astra, 46 for Kimi and 32 for DeepSeek. Those results are not proof of equivalent capability or evidence about how the models were trained.
What should AI customers do now?
The source recommends checking model-data provenance, testing models on relevant work, keeping a fallback and avoiding unofficial API resellers. The report does not indicate an immediate change to consumer apps.
Source: ThorstenMeyerAI.com
