🔍 Read the full analysis: The GPT-6 Astra Question Behind China’s New 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 recently released models. Reuters reported that the three companies did not immediately respond, and public details about the alleged training, data volumes and access methods remain unavailable.
The Information reported October 28 that Chinese AI labs DeepSeek and Moonshot AI relied heavily on answers from OpenAI’s GPT-6 Astra to train models released this month. Reuters relayed the allegation, naming Moonshot’s Kimi K3.1 Pro and DeepSeek’s V4.1 Flash update; none of the three companies had confirmed or denied it in Reuters’ account.
The Information’s report, by Jing Yang, Juro Osawa and Qianer Liu, cited people with direct knowledge, according to Reuters. Its public opening described Chinese AI labs as a growing customer group seeking to extract capabilities from Astra, which The Information called OpenAI’s most powerful released model. The rest of the article is paywalled, limiting what can be independently assessed from its public text.
Reuters said the models allegedly benefiting from Astra outputs were Kimi K3.1 Pro, released October 23, and DeepSeek’s V4.1-Flash-1017, launched October 17. Reuters reported that OpenAI, DeepSeek and Moonshot did not immediately respond to requests for comment. The source material provides no public figures for how many outputs were used, which prompts were involved, or how access to Astra was obtained.
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 disclose training data. Those statements do not independently confirm or disprove the allegation. A separate report by South Korea’s SBS, as summarized in the source material, quoted an OpenAI spokesperson saying the company takes the issue seriously; that account could not be checked against The Information’s paywalled report.
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 reported use is substantiated, the case could put a spotlight on the boundary between ordinary model distillation and using a rival’s service to build a competing product. Distillation trains a model to imitate answers from another model. It is a standard technique, including when a lab uses its own larger model to teach a smaller one. The dispute here is whether another provider’s outputs were used at scale and in a way that violated its terms.
OpenAI’s terms, according to the source material, prohibit using its outputs to build competing models. That raises a potential contract question; claims of fake or resold accounts, if established, could raise separate questions about how access was obtained. The available reporting does not establish that either lab broke a law or breached a contract, and legal conclusions would depend on facts not yet public.
The commercial stakes are also relevant. 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 scores were 57, 46 and 32 respectively. These figures suggest lower listed prices alongside lower scores on that index; they do not establish that the models are equivalent or explain how any model was trained.
For businesses choosing models, the report is a reminder to ask vendors about training-data provenance and to test systems on their own workloads. For individual users, no immediate change to the named apps is established by this report. Buyers should be cautious about unofficial API resellers, but the source material does not show that users of the official services are affected.
Products Worth Considering
DJI Mic 3 Wireless Microphone System (2 Transmitters, 1 Receiver, Charging Case)
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
As an affiliate, we earn on qualifying purchases.
Earlier Scrutiny of Model Distillation
The allegation follows earlier disputes over attempts to extract model capabilities. The source material says OpenAI linked a July campaign targeting hidden reasoning to people associated with Moonshot AI in September. 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 attributed claims and responses, not findings that settle the current report.
The source material also describes a U.S. government advisory as recognizing distillation as a legitimate and useful research technique. The distinction at issue is not simply whether one model learned from another’s outputs, but whose model supplied those outputs, the scale and method of collection, and whether the provider’s rules were breached.
There is a technical reason providers may guard detailed reasoning traces: training on fuller explanations can transfer more than final answers. The source 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 independent evidence about Astra or the Chinese models named in the report.
Evidence Still Missing From Public View
The central claim remains unverified publicly. The Information relied on people with direct knowledge, as Reuters reported, but the accessible material does not identify them or provide documentation. The companies’ lack of immediate responses in Reuters’ report is not confirmation of the allegation.
Key details remain unknown: the quantity and type of Astra outputs allegedly used, whether hidden reasoning was obtained, how accounts were created or supplied, and how much any such data contributed to the released models. The source material also does not provide a public technical comparison that could establish training overlap. Without those details, the extent and mechanism of the alleged distillation cannot be assessed.
It is also unclear whether OpenAI, DeepSeek or Moonshot will issue fuller statements, or whether any provider or regulator will publish evidence or take formal action. The reported conduct should not be treated as a proven legal violation.
Company Responses and Evidence to Watch
The next developments to watch are direct responses from OpenAI, DeepSeek and Moonshot, along with any further reporting that supplies evidence behind the account cited by The Information. Relevant details would include the scale of output collection, the access route, and whether the labs dispute the characterization that the models relied heavily on Astra.
Any formal complaint, investigation or policy action would clarify whether the dispute moves beyond reporting and contractual concerns. Until then, model buyers can review providers’ disclosures, check service terms and maintain alternatives suited to their own tasks. Those steps are practical precautions, not evidence that the allegation has been proven.
Key Questions
Which models did the report name?
The Information’s claim, as summarized by Reuters, concerned Moonshot AI’s Kimi K3.1 Pro and DeepSeek’s V4.1-Flash-1017 update.
Have the companies confirmed the reported use of Astra?
No. Reuters reported that OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment. The source material includes no direct confirmation from any of them.
Is distillation itself improper?
Not inherently. Distillation is a common method of training one model using another model’s outputs. The reported concern is whether a rival’s outputs were used at scale and in ways that may breach service terms or involve improper access.
Does the report prove a contract breach or illegal conduct?
No. The public details described here do not establish how the outputs were obtained or whether any rules were broken. The allegation remains unconfirmed, and any legal assessment would depend on further evidence.
Do users need to change anything now?
The source material identifies no immediate change to the named apps for users. Businesses may wish to review vendor disclosures and test alternatives; users should be cautious with unofficial API resellers.
Source: ThorstenMeyerAI.com
