🔍 Read the full analysis: China’s Newest AI Models And GPT-6 Astra: Separating Reported Findings From Unknowns on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
The Information reported on October 28 that DeepSeek and Moonshot AI relied heavily on OpenAI’s GPT-6 Astra outputs when developing recent models; Reuters relayed the claim. The companies have not publicly confirmed it, and the reporting available publicly does not specify the scale or method. Distillation is a common training technique, but using a rival’s outputs could breach OpenAI’s terms, depending on how access and training were handled.
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 recent models, a claim Reuters summarized within hours. The allegation could raise questions about model-training practices and OpenAI’s terms, but the publicly available reporting provides no figures or method, and none of the three companies had confirmed or denied it in comments to Reuters.
The models named in the report are Moonshot’s Kimi K3.1 Pro, released October 23, and DeepSeek-V4.1-Flash-1017, launched October 17. Reuters said The Information reported that both “benefitted from distilling Astra.” The Information’s article, by Jing Yang, Juro Osawa and Qianer Liu, was published at 6 p.m. Pacific time on October 28. Its publicly visible opening described Chinese AI labs as a growing customer base for Astra; the detailed article is paywalled.
Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to requests for comment. The report relies on people described as having direct knowledge, but their identities and the evidence supporting the claim are not available in the public account. It does not say how many Astra outputs were used, how they were obtained, or how much they affected the released models.
Moonshot’s launch post attributed Kimi K3.1 Pro’s gains to larger post-training compute and large-scale reinforcement learning. DeepSeek’s update note did not describe its training data. Those statements do not independently resolve the allegation: they offer limited public information about development, rather than a direct account of 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 Model Training Methods Matter
If substantiated, the report would put a commercial and policy question at the center of competition between AI developers: whether a company can use a rival’s outputs to train a competing model, and under what conditions. Distillation itself is a standard technique, including for making smaller, less expensive models. The potential dispute is about whose model served as the teacher, the scale of any copying, and whether access complied with the provider’s terms.
The source material lists prices of $10 per million input tokens and $50 per million output tokens for Astra, 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 respective scores of 57, 46 and 32 on the same independent index. Those figures suggest a price and performance gap in that comparison, not equivalent capability. The report does not establish that any price or score difference resulted from distillation.
For organizations choosing models, the practical issue is provenance and reliability, not just benchmark rankings or price. Buyers may want to ask providers how models were developed, assess them on their own tasks, and retain alternatives. The source article also advises users to avoid cheap, unofficial API resellers; the report itself does not establish that any particular reseller was involved.
Popular right now
Breville BES870XL Barista Express Espresso Machine, Brushed Stainless Steel
Vitamix 5200 Professional-Grade Blender, 64 oz. Container, Black
Dyson Car+Boat Cordless Handheld Vacuum
As an affiliate, we earn on qualifying purchases.
Distillation and Earlier Allegations
In distillation, a smaller “student” model is trained using answers generated by a larger “teacher” model. Researchers and companies use the approach to transfer capabilities or produce models that cost less to run. A U.S. government advisory cited in the source material calls distillation “a legitimate and useful technique in AI research.” That does not settle whether a specific use of another company’s outputs violates its terms or other rules.
The source article describes earlier allegations and official responses as background. It says OpenAI, in September, connected a “core cluster” of a July campaign to extract hidden reasoning from its models to people associated with Moonshot. It also says U.S. security agencies named DeepSeek and Moonshot in an advisory, while China’s commerce ministry called the accusations baseless. These accounts are not independent confirmation of the new report’s claim about training Kimi K3.1 Pro or DeepSeek-V4.1-Flash-1017.
The article also cites Korean broadcaster SBS, which said Chinese labs had previously focused on Anthropic’s Claude and shifted toward OpenAI after Anthropic tightened defenses. SBS reportedly quoted an OpenAI spokesperson saying the company takes the issue seriously. The source writer says these details could not be checked against The Information’s full report, so they should be treated as secondary reporting, not independently verified findings.
“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 Claim
The available account does not reveal the number of Astra outputs involved, the prompts used, or the route by which any outputs were obtained. It also does not establish whether the models’ training relied heavily on those outputs or whether they were one input among several. The underlying article is paywalled, and the reporting described here does not include documents or a public technical analysis that would let readers assess the allegation independently.
It remains unknown whether any access involved fake or resold accounts, whether OpenAI’s terms were breached, or whether any law was violated. The source material frames the likely dispute as a possible contract issue, with possible concerns about fraud depending on how access was obtained; those are not legal findings. No public response from the three companies addressing the report’s central claim was included in the Reuters account.
The performance and price figures in the source are presented without enough detail here to assess the index, testing conditions or comparability of the products. They should not be taken as proof of how either model was trained or as a complete measure of their usefulness.
Responses and Verification Ahead
The next useful development would be a direct response from OpenAI, DeepSeek or Moonshot, or publication of supporting evidence that clarifies the scale and source of the alleged outputs. The companies may also explain their account controls, training processes or terms, though no such disclosure or timetable was reported in the available material.
Until then, the claim remains a reported allegation rather than an independently established account. Organizations evaluating these models can review provider statements and model documentation, test performance on their own workloads, and keep a fallback option. Consumers should distinguish official services from unofficial API resellers, while avoiding assumptions that the report changes how the apps work today.
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’s Kimi K3.1 Pro. Reuters relayed the claim; the companies had not confirmed or denied it in the account provided.
What is distillation in AI?
Distillation trains a smaller model using outputs from a larger model, with the aim of transferring some of its capabilities. It is a common technique. Whether a particular use violates a provider’s terms depends on the circumstances and the applicable agreement.
Has OpenAI confirmed that the labs used Astra?
No confirmation is included in the available reporting. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment.
Does the report show that laws were broken?
No. The report, as publicly summarized, does not establish a legal violation. Possible issues could depend on OpenAI’s terms and how access was obtained; the available material does not settle those questions.
What should AI buyers do now?
They can ask providers about model provenance and training documentation, test models on their own tasks, and maintain alternatives. The report does not establish that the named models are unusable or that their current services have changed.
Source: ThorstenMeyerAI.com
