🔍 Read the full analysis: China’s Newest AI Models: Examining The Report On GPT-6 Astra on ThorstenMeyerAI.com
Get audio and creator 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 OpenAI’s GPT-6 Astra outputs to train recent models, citing people with direct knowledge. The report has not been confirmed by the companies, and it offers no public figures or details about how the outputs were obtained.
The Information reported on 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’s V4.1 Flash update and Moonshot’s Kimi K3.1 Pro. Reuters summarized the report, which cites people with direct knowledge, but none of the three companies had confirmed or denied the claim in its account.
The models named are DeepSeek-V4.1-Flash-1017, released October 17, and Kimi K3.1 Pro, released October 23, according to the source article’s account of The Information and Reuters. The Information’s public opening described Astra as a growing source of demand among Chinese AI labs seeking to extract its capabilities; the article’s detailed reporting is behind a paywall. Reuters reported that both models “relied heavily” on Astra outputs, but the public reporting cited here provides no volume estimates, technical method or account-level evidence.
Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to requests for comment. A separate summary by South Korea’s SBS, as described in the source article, said OpenAI takes the issue seriously and continues to work on it. That summary also attributed to sources a claim that some Chinese labs shifted attention from Anthropic’s Claude to OpenAI after Anthropic tightened defenses. Those details could not be checked against The Information’s full report and should be treated as separately attributed, unverified reporting.
The models’ own release materials provide a different account of their development, without addressing the allegation directly. Moonshot’s launch post credited Kimi K3.1 Pro’s gains to larger post-training compute and large-scale reinforcement learning. DeepSeek’s update note, as described in the source article, did not discuss training data. Neither statement settles 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
The Stakes for AI Model Training
The allegation centers on a distinction between a common training technique and how it is used. Distillation involves training a smaller model on outputs from a larger “teacher” model. AI labs use it to make less costly models, including by training on answers from their own systems. A U.S. government advisory cited in the source article calls distillation a legitimate and useful research technique.
The dispute, if the report is substantiated, is whether a rival’s model outputs were collected at large scale and used to build competing systems in a way that breached access terms or involved deceptive account use. OpenAI’s terms, according to the source article, bar customers from using its outputs to build competing models. That makes the matter potentially a contract and access-compliance issue; whether any law was broken is not established by the reporting presented here.
The reported economics also help explain the interest in the claim, but do not prove anything about training. The source article 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. It also gives scores of 57, 46 and 32 on one independent index, respectively. The article does not name the index or provide its methodology in the supplied material. Those figures suggest lower listed prices for the two Chinese models, while the scores do not show identical performance.
For organizations choosing AI services, the immediate practical issue is model provenance: how a model was trained, what data or outputs it used, and whether its development complied with relevant terms. The source article advises businesses to conduct due diligence, retain a fallback model and test systems on their own tasks. For users, it says there is no reported change to the apps today, while warning against unofficial, low-cost API resellers whose access arrangements may be unclear.
Products Worth Considering
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Earlier Concerns About Model Extraction
The latest report follows earlier public concerns about attempts to obtain AI models’ capabilities through large-scale prompting. The source article says OpenAI in September linked what it called a “core cluster” of a July campaign to extract hidden reasoning from its models to people 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 prior claims and responses; they do not independently verify the new report about Astra.
The source article also refers to a February 2026 Anthropic account alleging that DeepSeek, Moonshot and MiniMax generated more than 16 million exchanges with Claude. The provided material cuts off before giving the full details of that episode, so its scope and findings cannot be assessed here. The history offers relevant context for why model providers monitor unusual access patterns, but the current Astra allegation remains distinct and has not been substantiated publicly in the material provided.
Providers may limit access to detailed reasoning traces because those traces can make a model’s problem-solving behavior easier to imitate. The source article 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 company-run test illustrates why detailed outputs may be useful for training; it does not establish that DeepSeek or Moonshot obtained or used Astra reasoning traces.
Evidence Behind the Astra Claim
The public account leaves major questions unanswered: how many Astra outputs were used, what prompts or accounts obtained them, and whether the outputs included hidden reasoning. The report, as summarized by Reuters and the source article, relies on unnamed sources with direct knowledge. The underlying detailed story is paywalled, and no supporting data or technical analysis is included in the supplied material.
It is also unclear whether the companies will respond, whether OpenAI has provided evidence to regulators or other authorities, or whether any investigation is under way. The report does not establish a legal violation. The source article characterizes the potential dispute mainly as a question of contract terms and possible fraud in obtaining access, not as a proven criminal case.
Company Responses and Verification
The next developments to watch are responses from OpenAI, DeepSeek and Moonshot, along with any additional reporting that discloses evidence, methods or scale. The companies’ public training disclosures may also clarify what they say contributed to the models’ capabilities, though their existing release notes do not resolve the allegation.
Until more evidence is made public, businesses evaluating these systems can review vendor statements and access terms, test model performance on their own work and maintain alternatives. Users should distinguish official services from resellers. The central unanswered issue is whether the reported use of Astra outputs occurred as described and, if so, how access was obtained and what role those outputs played in the released models.
Key Questions
Which models are named in the report?
The report names Moonshot AI’s Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, released October 17. The allegation is that they relied heavily on outputs from OpenAI’s GPT-6 Astra during training.
Have the companies confirmed the report?
No. Reuters reported that OpenAI, DeepSeek and Moonshot did not immediately respond to requests for comment. The supplied material does not include a direct confirmation or denial from any of them.
Is distillation itself improper?
Not necessarily. Distillation is a standard technique for training a smaller model on a larger model’s outputs. The concern raised by the report is the alleged use of a competitor’s outputs at scale and whether access or use breached the provider’s terms.
Does the report prove that a law was broken?
No. The public information described here does not establish a legal violation. The source article says OpenAI’s terms prohibit using its outputs to build competing models, making contract compliance and access methods central questions. Any legal conclusion would require more evidence.
Source: ThorstenMeyerAI.com
