🔍 Read the full analysis: A Closer Look At China’s New AI Models And The GPT-6 Astra Report on ThorstenMeyerAI.com
Get smart everyday buys 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 outputs from OpenAI’s GPT-6 Astra to train recent models. Reuters repeated the claim, citing the report; the companies had not responded to Reuters at publication, and public evidence on the methods and scale remains limited.
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-1017. Reuters summarized the allegation, but the public record does not establish how the outputs were obtained or how extensively they shaped either model, and the companies had not responded to Reuters at the time of its report.
The Information’s article, by Jing Yang, Juro Osawa and Qianer Liu, was published at 6 p.m. Pacific time on Oct. 28. Its public opening described Astra as OpenAI’s most powerful released model and said demand from Chinese AI labs seeking to extract its capabilities was growing. The rest of the article is paywalled. Reuters reported that the labs’ latest models had benefited from distilling Astra, citing people with direct knowledge, but did not provide public evidence or detailed methods.
The models named were Moonshot’s Kimi K3.1 Pro, released Oct. 23, and DeepSeek’s V4.1-Flash-1017, released Oct. 17. Moonshot’s launch post attributed Kimi’s gains to substantially more post-training computing and large-scale reinforcement learning. DeepSeek’s update note did not describe its training data. Those statements do not independently confirm or disprove the allegation.
Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to requests for comment. South Korea’s SBS separately summarized the report on Oct. 29 and quoted an OpenAI spokesperson as saying the company takes the issue seriously and continues working on it. The original reporting and that statement have not been independently detailed in the public material provided here.
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 Access Matters
The allegation matters because model distillation can let developers reproduce some capabilities of a more expensive system at a lower operating cost. If a lab uses a rival’s outputs at scale, the dispute can affect competition, access controls and the terms governing AI services. The report, however, does not establish a breach or show how much the named models’ performance came from Astra rather than their developers’ own training work.
The source article compared listed API prices: it gave Astra’s rates as $10 per million input tokens and $50 per million output tokens, Kimi K3.1 Pro’s as $2 and $10, and DeepSeek’s as $0.30 and $1.20 during peak hours. It also cited scores of 57, 46 and 32, respectively, on the same independent index. Those figures are the article’s comparison, not proof that the models are interchangeable; the price and score differences point to meaningful distinctions in cost and measured performance.
For organizations choosing models, the episode is a reminder to assess data provenance, provider terms and performance on their own tasks, rather than rely on price or a single benchmark. For ordinary users, the report does not establish an immediate change to the named services. The source article advises avoiding unofficial, low-cost API resellers, whose access arrangements may be unclear.
Popular right now
Ninja DualZone 6-in-1 Air Fryer, 8-Qt, Grey, DZ801
Vitamix 5200 Professional-Grade Blender, 64 oz. Container, Black
Bose QuietComfort Wireless Noise Cancelling Headphones, Sandstone
As an affiliate, we earn on qualifying purchases.
How Distillation Became the Dispute
Distillation is a standard training approach: a developer uses answers from a larger “teacher” model to train a smaller “student” model to produce similar responses. The U.S. government advisory cited in the source material calls it a legitimate and useful AI research technique. The controversy is not the method by itself, but whether a developer uses a competitor’s model at industrial scale in ways barred by the provider’s terms or enabled by deceptive account access.
The source article says OpenAI’s terms prohibit using its outputs to build competing models. That makes a potential dispute partly a question of contract terms and how access was obtained; whether any law or contract was violated cannot be determined from the reporting available. The article also describes earlier concerns: in September, OpenAI associated a July campaign to extract hidden reasoning with individuals it linked to Moonshot AI, while U.S. security agencies named both labs in an advisory. China’s commerce ministry called the accusations baseless, according to the source material. Those earlier claims are separate from proof of the specific Astra-training allegation.
The Information’s public excerpt said Astra had attracted strong demand since its early-September release. Beyond that, the article’s detailed evidence is behind a paywall. The source article also described a prior Anthropic experiment in which models trained on full reasoning transcripts performed better on a coding benchmark than models trained on summaries. That is Anthropic’s own test, not evidence that DeepSeek or Moonshot obtained Astra’s hidden reasoning.
“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 Allegation
The public reporting supplied here does not disclose how many Astra outputs were used, which prompts or accounts were involved, or whether access came through direct subscriptions, resellers or another route. It also does not quantify Astra’s contribution to either model’s training or performance. The Information cited people with direct knowledge, but its detailed account is paywalled, leaving outside readers unable to assess the full evidence from the available text.
No public confirmation or denial from OpenAI, DeepSeek or Moonshot was included in Reuters’ report. The source material says China’s commerce ministry rejected related accusations as baseless, but that statement does not resolve the specific claim about Astra. It also remains unclear whether any provider has made a formal contractual complaint or whether regulators are examining this particular report.
Claims that Chinese labs previously focused on Anthropic’s Claude and shifted toward OpenAI after stronger defenses were reported by SBS as part of its summary; the source article says those details could not be checked against the original. They should be treated as attributed reporting, not established fact.
Responses and Evidence to Watch
The immediate next step is whether OpenAI, DeepSeek or Moonshot provides a substantive response, including details on account access, training data or any review of the claims. Further reporting could clarify the sources’ basis, the scale of model-output use and whether the companies dispute the account. No timetable for a company statement or investigation is given in the supplied material.
For users and business customers, the practical course is to watch for official statements and any changes to API terms or access controls, while testing models directly for their own needs. The report alone does not establish that Kimi K3.1 Pro or DeepSeek V4.1 Flash is unusable, nor does it prove that either model’s performance came primarily from Astra. Any assessment should distinguish the allegation from the confirmed release dates and the published pricing and benchmark comparisons.
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 Kimi K3.1 Pro and DeepSeek-V4.1-Flash-1017. Reuters summarized the claim, which remains an allegation in the public material.
Have the companies confirmed the report?
No confirmation is included. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment. The available material does not provide a public denial either.
Is model distillation itself improper?
Not inherently. Distillation is a commonly used training method, and a U.S. government advisory quoted in the source material calls it a legitimate research technique. The dispute concerns possible use of a competitor’s outputs at scale and whether that would violate service terms or involve improper access.
What evidence has been made public?
The Information cited people with direct knowledge, but its detailed article is paywalled. The public material does not provide usage figures, account details, training records or a measured contribution from Astra to either model.
Does the report change anything for users now?
The report does not establish an immediate change to the apps or models. Users and organizations can follow official statements and use caution with unofficial API resellers, but the allegation alone does not show that the named models are unavailable or that their outputs are unreliable.
Source: ThorstenMeyerAI.com
