🔍 Read the full analysis: China’s Latest AI Models And GPT-6 Astra: Findings, Questions, And Limits on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
The Information reported on Oct. 28 that DeepSeek and Moonshot AI relied heavily on OpenAI’s GPT-6 Astra outputs to train recent models. Reuters repeated the claim, but the companies had not responded to Reuters, and the public record does not identify the methods or scale. Distillation is a standard technique; the dispute concerns whether a rival’s model outputs were collected in ways that breach service terms.
The Information reported on Oct. 28 that Chinese AI companies 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 report names Kimi K3.1 Pro and DeepSeek’s V4.1 Flash update, but the public account offers no disclosed evidence about the amount of data or how it was obtained, and the three companies had not responded to Reuters.
The Information’s article, by Jing Yang, Juro Osawa and Qianer Liu, was published at 6 p.m. Pacific time on Oct. 28. Its publicly visible opening describes Astra as OpenAI’s most powerful released model and says Chinese AI labs are among a growing customer base seeking to extract its capabilities. The remainder of the article is behind a paywall. Reuters reported that the labs’ latest models “relied heavily” on Astra outputs, citing The Information.
The models identified in subsequent coverage are Moonshot’s Kimi K3.1 Pro, released Oct. 23, and DeepSeek’s V4.1-Flash-1017, launched Oct. 17. Moonshot’s release post attributed Kimi’s gains to substantially more post-training compute and large-scale reinforcement learning. DeepSeek’s update note, as described in the source material, did not detail training data. Neither company has publicly confirmed the reported use of Astra outputs.
Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its request for comment. SBS, a South Korean broadcaster, separately summarized the report on Oct. 29 and quoted an OpenAI spokesperson as saying the company takes the issue very seriously and continues working on it. That quotation and other details in SBS’s account could not be checked against the paywalled original, so they do not independently establish how the models were trained.
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 substantiated, the report could affect how AI providers enforce access terms and how customers assess the origins of models they deploy. OpenAI’s terms prohibit using its outputs to build competing models, according to the source material. Whether the alleged activity violated those terms would depend on facts not publicly available, including how the outputs were obtained, what restrictions applied, and what use was made of them.
The commercial stakes are also visible in the listed prices and benchmark scores. The source gives Astra prices of $10 per million input tokens and $50 per million output tokens; Kimi K3.1 Pro at $2 and $10; and DeepSeek’s model at $0.30 and $1.20 during peak hours. On the same independent index, the models score 57, 46 and 32, respectively. Those figures suggest cheaper alternatives with lower scores on that index, not equivalent performance; the source does not provide enough information here to generalize the scores to every task.
For companies choosing models, the report is a reason to ask providers about training-data provenance, maintain fallback options and test systems on their own workloads. It does not, by itself, show that users need to change apps or that the named models are unusable. The source advises users to avoid cheap, unofficial API resellers, but does not document that any particular reseller supplied access in this case.
Products Worth Considering
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
DJI Mic 3 Wireless Microphone System (2 Transmitters, 1 Receiver, Charging Case)
Logitech MX Mechanical Wireless Illuminated Keyboard Tactile (Graphite)
As an affiliate, we earn on qualifying purchases.
Distillation and the Earlier Dispute
Distillation is a common training method: a smaller model learns to imitate outputs from a larger “teacher” model. AI labs use it to produce more efficient versions of their own systems, and a U.S. government advisory described the technique as legitimate and useful in AI research. The method itself is not evidence of wrongdoing. The dispute here is whether a rival’s outputs were used at scale and whether access complied with the provider’s rules.
The report follows earlier allegations about attempts to extract capabilities from leading AI systems. The source material says that in September OpenAI linked a “core cluster” of a July campaign targeting hidden reasoning to people it 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. Those are separate allegations and responses; they do not independently verify The Information’s account about Astra or the named October models.
The source also recounts 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 reported test, not evidence about DeepSeek’s or Moonshot’s training. It illustrates why access to detailed model outputs may matter, while leaving the Astra allegation unproven.
“DeepSeek and Moonshot “relied heavily on outputs” from GPT-6 Astra to train their latest models.”
— The Information, as summarized by Reuters
Evidence and Access Methods Remain Unknown
The public reporting does not disclose how many Astra outputs were used, which prompts or accounts were involved, whether hidden reasoning was obtained, or how the data shaped either model. The Information cites people with direct knowledge, but its detailed account is paywalled; Reuters’s report summarizes the allegation rather than supplying underlying records. No company had confirmed or denied the central claim in Reuters’s account.
It is also unclear whether any access involved fake or resold accounts, a possibility discussed as a general concern in the source material rather than established in this case. The available material does not show whether OpenAI has made a formal legal claim, whether regulators are investigating this specific report, or whether the named labs have changed their training practices. The claim should therefore remain attributed, not treated as a finding of misconduct.
Company Responses and Verification
The next developments to watch are direct responses from OpenAI, DeepSeek and Moonshot, and any further reporting that provides evidence about the scale and method of the alleged training. The source material gives no announced investigation, deadline or scheduled company statement, so none can be assumed.
For now, organizations evaluating these models can ask vendors about training-data sources and service compliance, compare performance on their own tasks, and retain alternatives where appropriate. Those steps address provenance uncertainty without treating the reported allegation as established fact. Any legal or contractual conclusions will depend on evidence and the terms applicable to the access in question.
Key Questions
What did the report allege?
The Information reported that DeepSeek and Moonshot relied heavily on GPT-6 Astra outputs to train recent models, including Kimi K3.1 Pro and DeepSeek V4.1-Flash-1017. The public account does not disclose the alleged volume or method.
Have the companies confirmed the claim?
No. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its requests for comment. The report remains an attributed allegation, not a publicly confirmed account.
Is distillation itself prohibited?
Distillation is a standard AI training technique and has legitimate uses. The potential dispute is whether a company used a competitor’s outputs in a way barred by service terms or obtained them through improper access. The available report does not establish those details.
Do the reported models match Astra’s performance?
The supplied source lists scores of 57 for Astra, 46 for Kimi K3.1 Pro and 32 for DeepSeek’s model on one independent index. It also lists lower prices for the Chinese models. These figures do not establish equal performance across tasks, and the index’s broader limits are not detailed here.
Source: ThorstenMeyerAI.com
