🔍 Read the full analysis: How Strong Is The Case That China’s AI Models Learned From GPT-6 Astra? on ThorstenMeyerAI.com
Get privacy and security gear 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 outputs from OpenAI’s GPT-6 Astra to train recent models. Reuters said the three companies did not immediately respond, and the public record offers no figures or method showing how the outputs were obtained. The allegation remains unconfirmed; distillation is a standard technique, but using a rival’s model may violate its 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 models released this month. Reuters summarized the allegation, but the public reporting provides no figures or account of how the outputs were obtained, and none of the three companies had responded to Reuters at the time of its report.
The models named are Moonshot’s Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, released October 17. The Information’s publicly visible opening described Chinese AI labs as a growing customer base for Astra; the detailed account is behind a paywall. Reuters reported that the labs’ latest models “relied heavily” on Astra outputs and said they “benefitted from distilling” the model.
The report cites people with direct knowledge, according to Reuters’ account. It does not publicly specify how many prompts were sent, which Astra outputs were used, whether access came through ordinary accounts or intermediaries, or how much those outputs contributed to either model. There is no public confirmation from DeepSeek, Moonshot or OpenAI in the source material.
The labs’ own release materials point to other training factors, but do not settle the allegation. Moonshot credited Kimi K3.1 Pro’s gains to larger post-training compute and large-scale reinforcement learning. DeepSeek’s update note did not discuss training data. Those statements do not independently establish whether Astra outputs were used.
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 Training Provenance Matters
If the report is substantiated, the case could test how model providers enforce rules against using their outputs to build competing systems. OpenAI’s terms prohibit using its outputs to develop competing models, according to the source material. That makes the central issue partly contractual; whether access involved deception, account resale or other conduct is not established by the public reporting.
The commercial stakes are visible in the reported price and performance comparisons. The 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. On the same independent index, it says the models scored 57, 46 and 32, respectively. These figures suggest lower prices alongside lower scores on that index; they do not establish equivalent performance or explain whether any gap is caused by training data.
For businesses evaluating AI systems, the story is a reminder to ask providers about training-data provenance and licensing, while testing models on their own workloads. For consumers, the source reports no immediate change to the named apps. It advises caution around cheap, unofficial API resellers, although it provides no evidence that a particular reseller supplied access in this case.
Products Worth Considering
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
Logitech MX Mechanical Wireless Illuminated Keyboard Tactile (Graphite)
Samsung Galaxy Tab S10 FE 128GB Wi-Fi Android Tablet, Blue
As an affiliate, we earn on qualifying purchases.
Distillation and Earlier Allegations
Distillation is a standard model-training technique: a smaller “student” model is trained on outputs from a larger “teacher” model. The U.S. government advisory cited in the source calls it a legitimate and useful research method. The dispute is not over the technique alone, but whether a lab used another company’s model at scale in a way that breaches terms or access rules.
The report follows earlier concerns raised by AI companies and U.S. agencies. In September, OpenAI said a “core cluster” of a July campaign to extract hidden reasoning from its models was tied to individuals associated with Moonshot AI. The source also says U.S. security agencies named both DeepSeek and Moonshot in an advisory, while China’s commerce ministry called the accusations baseless. These are attributed statements, not independent proof of the current claim.
South Korean broadcaster SBS reported October 29 that sources said Chinese labs had mainly targeted Anthropic’s Claude before shifting 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 checked against the paywalled Information report in the supplied material, so they remain secondary reporting rather than independently verified details.
“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 Distillation Claim
The core allegation has not been publicly substantiated with prompt counts, training records, account details or a described method. The Information cited people with direct knowledge, but its full report is paywalled; Reuters’ summary does not provide the evidence needed to assess the scale or precise role of Astra outputs.
It is also unclear whether Astra was accessed directly, through resellers, or by another route, and whether any access violated OpenAI’s terms. No public response from OpenAI, DeepSeek or Moonshot is included in the supplied Reuters account. The reported links to earlier activity and claims about a shift from Claude are separate claims and should not be treated as proof of this specific episode.
Finally, the reported benchmark scores and prices do not reveal how either Chinese model was trained. Without further documentation, readers cannot determine whether the models’ performance reflects Astra distillation, the labs’ stated training methods, other data, or some combination.
Company Responses and Further Evidence
The next developments to watch are responses from OpenAI, DeepSeek and Moonshot, along with any additional reporting that identifies the scale, access route or training evidence. The supplied material gives no scheduled company statements, investigation findings or regulatory milestone.
For now, the report remains an allegation based on sources described by The Information and summarized by Reuters. Clearer conclusions would require direct company comment or verifiable details about the prompts, accounts and training pipelines involved. Until then, the distinction between normal distillation and an alleged breach of a provider’s rules remains central.
Key Questions
What is the allegation about DeepSeek and Moonshot?
The Information reported that the labs relied heavily on GPT-6 Astra outputs to train DeepSeek-V4.1-Flash-1017 and Moonshot’s Kimi K3.1 Pro. The companies had not confirmed or denied the claim in Reuters’ account.
Does using distillation prove a company broke the rules?
No. Distillation is a common training technique. The disputed questions are whether a rival’s outputs were used, whether access or use violated the provider’s terms, and what evidence supports the allegation.
What evidence has been made public?
The supplied reporting attributes the claim to people with direct knowledge, but gives no public figures, prompt records or account details. The full Information report is paywalled, and the companies’ responses were not available in Reuters’ account.
What does this mean for AI users right now?
The source reports no immediate change to the named apps. Businesses may want to ask AI vendors about model provenance and test performance on their own tasks; users should be cautious with unofficial API resellers.
Source: ThorstenMeyerAI.com
