The Information Says DeepSeek and Moonshot Trained on GPT-6 Astra. Nobody Has Confirmed It, and OpenAI's Anti-Distillation Rule Starts November 20.
The Information reported on October 28, in a paywalled piece relayed by The Next Web and Investing.com, that DeepSeek and Moonshot used GPT-6 Astra outputs to train Kimi K3.1 Pro and DeepSeek V4.1 Flash 1017, two of October's top open-weight scorers. No statement from OpenAI, DeepSeek, or Moonshot has been found. The one verified fact is the clock: OpenAI's expanded usage policies on distillation and intermediaries take effect November 20, 2026. We read the relays and OpenAI's policy page.
By T.W. Ghost
The Information reported today, October 28, 2026, that DeepSeek and Moonshot used outputs from OpenAI's GPT-6 Astra to train their latest models, naming Kimi K3.1 Pro and DeepSeek V4.1 Flash 1017. The piece is paywalled and anonymously sourced; we read it through The Next Web's and Investing.com's relays. Neither OpenAI, DeepSeek, nor Moonshot has confirmed, denied, or commented in anything we could find.
That is the whole evidentiary record, and it is an allegation. What is verified sits next to it: OpenAI's expanded usage policies, which add an anti-distillation rule and an approval requirement for intermediaries, take effect November 20, 2026. We covered those when they were published. The report and the rule are related by subject and not by evidence; the policy was written before the story ran, and the story does not make the policy any more or less true. Here is what held up.
What Is Alleged, and What Is Verified
| Item | Status |
|---|---|
| DeepSeek and Moonshot trained on GPT-6 Astra outputs; Kimi K3.1 Pro and DeepSeek V4.1 Flash 1017 named | Press, paywalled and anonymously sourced (The Information, via relays) |
| Any confirmation or denial from OpenAI, DeepSeek, or Moonshot | None found |
| OpenAI usage policies: anti-distillation and intermediary approval, effective November 20, 2026 | Verified (OpenAI policy page) |
| The named models' scores | Verified elsewhere: Kimi K3.1 Pro 55 on v4.3.2 and 46 on v5; Flash 1017 45 on v4.3.2 and 32 on v5, per Artificial Analysis |
| A commentator's estimate that about half of some lab's training signal is distilled | Commentary, not evidence |
Anonymous and paywalled means label it, not repeat it. We print the claim because readers will see it and because it names two models on our boards. We do not print it as fact, and nothing in the relays rises above the original's sourcing.
The clock is forward-looking. November 20 governs what OpenAI's customers may do with its outputs from that date. The named models already shipped. If the allegation is true, the rule arrives after the fact; if it is false, the rule still arrives.
Two of the month's best open scores are now under a question mark they did not have yesterday. That is the practical effect, and it is an effect on reputation, not on any measured number.
What To Do
Quote the allegation as an allegation with the outlet's name attached, and quote the policy date as the fact.
If you route through an intermediary or fine-tune on model outputs, read the November 20 terms now; they apply to you regardless of what The Information's sources saw.
The frontier is elsewhere: the scores stand until someone with a name says otherwise. Our model comparison has the current standings side by side.
Sources
- The Next Web: DeepSeek and Moonshot reportedly used GPT-6 Astra to train models
- Investing.com: DeepSeek, Moonshot used OpenAI's GPT-6 Astra to train latest models, The Information reports
- The Information: Astra model topic page
- OpenAI usage policies
*Not sure how a distillation allegation should change which open model you trust? Take the free 2-minute quiz and get matched. Then read the policy that is actually dated: OpenAI's usage policies for intermediaries.*