🔍 Read the full analysis: China’s Latest AI Models: Understanding The GPT-6 Astra Distillation Claims 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 Oct. 28 that DeepSeek and Moonshot AI relied heavily on OpenAI’s GPT-6 Astra outputs to train recently released models. Reuters said the three companies did not immediately respond; the public evidence does not establish how the outputs were obtained or how much they influenced training.
The Information reported Oct. 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 report, based on people with direct knowledge, has not been confirmed by the companies, and details about the alleged training or access methods remain unavailable publicly.
Reuters summarized The Information’s report within hours, identifying Moonshot’s Kimi K3.1 Pro and DeepSeek’s V4.1-Flash-1017 as models that benefited from distilling Astra. Kimi K3.1 Pro was released Oct. 23, while DeepSeek’s update launched Oct. 17. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to requests for comment.
The Information’s publicly available opening described Astra as OpenAI’s most powerful model and said Chinese AI labs were among a growing customer base seeking to extract its capabilities. The remainder of its report is paywalled. The public material does not give figures for how many outputs were used, the prompts or methods involved, or what portion of either model’s training came from Astra.
Moonshot’s launch post attributed Kimi K3.1 Pro’s gains to substantially greater post-training compute and large-scale reinforcement learning. DeepSeek’s update note did not describe its training data. Those statements do not independently confirm or disprove 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
The Stakes for AI Training
If the report is accurate, it highlights a growing dispute over how AI companies can use competitors’ model outputs to train new systems. Distillation—training a student model to imitate a teacher model’s answers—is a standard technique. The contested questions are whether the teacher is a rival’s service, whether output collection is conducted at scale, and whether access violates service terms or involves deceptive accounts.
The report also matters to customers weighing model cost against capability. The source material 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. On the same independent index cited there, the models score 57, 46 and 32, respectively. These figures suggest lower prices alongside lower scores on that index—not equivalent performance—and the source does not specify the index’s methodology or date.
For companies selecting models, the allegations are a reminder to ask vendors about training-data provenance, test systems on their own workloads and keep alternatives available. For individual users, no immediate change to the apps is established by this report. The source advises caution with low-cost, unofficial API resellers, whose access arrangements may be unclear.
Products Worth Considering
Samsung Galaxy Tab S10 FE 128GB Wi-Fi Android Tablet, Blue
LG UltraGear 32GX850A-B 32-Inch 4K UHD Glossy OLED Gaming Monitor
Bose Lifestyle Ultra Wireless Home Speaker
As an affiliate, we earn on qualifying purchases.
Distillation and Earlier Allegations
In distillation, a larger model generates responses that are used to train a smaller one to reproduce useful behavior. The U.S. government advisory cited in the source calls distillation a legitimate and useful technique in AI research. Its ordinary use does not, on its own, establish wrongdoing; the dispute described here concerns alleged use of a competitor’s outputs and the way access may have been obtained.
The report lands amid prior accusations. The source says that in September OpenAI linked a “core cluster” of a July campaign to extract hidden reasoning from its models to individuals associated with Moonshot AI. It also says U.S. security agencies named DeepSeek and Moonshot in an advisory, while China’s commerce ministry called the accusations baseless. These are claims and responses from the parties and agencies involved, not proof that the latest reported training occurred.
South Korean broadcaster SBS, in its Oct. 29 summary, reported further details attributed to sources: 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 seriously and continues working on it. The original report’s full sourcing is not publicly accessible, so these details cannot be checked against its paywalled text from the material available.
“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 Still Missing Publicly
The central claim remains unverified in the public record described here. The Information cited people with direct knowledge, but its detailed report is paywalled; Reuters said the three companies did not immediately respond. No public documentation in the source establishes the number of Astra outputs used, how they were collected, or their influence on either model’s final performance.
It is also unclear whether any account access was obtained through fake, shared or resold accounts, or whether the alleged activity breached a contract or law. OpenAI’s terms, according to the source, prohibit using its outputs to build competing models, but whether that term applies to the conduct alleged—and what remedies might follow—would depend on facts not publicly established. The claims do not by themselves establish a legal violation.
Finally, the models’ release notes point to other training methods, but those explanations do not settle whether Astra outputs were also used. The available information does not quantify the contribution of any one training source.
Company Responses and Further Evidence
The immediate next development would be a substantive response from OpenAI, DeepSeek or Moonshot, or additional reporting that makes the sourcing and alleged methods clearer. Reuters reported that the companies had not immediately responded; the source material provides no later confirmation or denial from them.
Readers should watch for evidence that distinguishes ordinary distillation from alleged use of a competitor’s service in breach of its terms, including information about account access, collection scale and training data. Until that emerges, the report should be treated as an allegation rather than a confirmed account of how the two models were built.
For organizations evaluating these systems, practical steps include checking vendor disclosures, testing performance on relevant tasks, reviewing API suppliers and maintaining a fallback model. The report does not establish that users need to change their current apps, and it gives no announced timetable for an investigation or company action.
Key Questions
What did The Information report?
It reported that DeepSeek and Moonshot AI relied heavily on GPT-6 Astra outputs to train recent models, naming DeepSeek V4.1-Flash-1017 and Moonshot Kimi K3.1 Pro. The claim is attributed to people with direct knowledge and has not been publicly confirmed by the companies.
What is AI distillation?
Distillation trains a smaller “student” model using responses from a larger “teacher” model, with the aim of reproducing some of its capabilities. It is a common research and development technique; questions about using a rival’s outputs or violating service terms are separate issues.
Do the reports prove that DeepSeek or Moonshot broke the law?
No. The public information described in the source does not establish how outputs were obtained or whether any law or contract was violated. OpenAI’s terms reportedly prohibit using its outputs to build competing models, but applying those terms to the alleged conduct would require more facts.
Have the companies confirmed the allegation?
No confirmation is included in the available source material. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment.
Do users need to change anything now?
The report does not establish an immediate change for users of the apps. Businesses can review vendor disclosures and API providers, test models on their own tasks and keep alternatives available while the claims remain unresolved.
Source: ThorstenMeyerAI.com
