THE ESSENTIALS IN 30 SECONDS

  • According to The Information, relayed by Reuters on October 28, both Chinese labs relied heavily on GPT-6 Astra outputs when training their newest models.
  • The models named are Moonshot's Kimi K3.1 Pro and DeepSeek's V4.1-Flash 1017, released on October 17.
  • OpenAI, DeepSeek and Moonshot AI did not immediately reply to Reuters; the claims have not been independently confirmed.

Two of China's best-known AI developers, DeepSeek and Moonshot AI, made heavy use of answers generated by OpenAI's GPT-6 Astra while training the models they shipped this month, according to a report by The Information published on October 28. Reuters relayed the story but did not say it had verified it, and the report rests on people described as having direct knowledge of the matter. The evidence behind the claim has not been made public.

What the report says

The Information singles out two releases. The first is Kimi K3.1 Pro, which Moonshot put out the week before the report. The second is DeepSeek's V4.1-Flash 1017, an updated model launched on October 17. In both cases, the outlet says, the developers benefited from distilling Astra, meaning they used the larger model's outputs as training material so that their own systems could pick up its behaviour.

OpenAI, DeepSeek and Moonshot AI did not respond right away when Reuters asked for comment. Until one of them does, or the underlying evidence surfaces, the account should be treated as an unconfirmed allegation.

A pattern of accusations

The report lands in an already tense context. In a post at the end of September, OpenAI said it had broken up a coordinated campaign, active from July, that tried to coax its models into revealing their protected reasoning. According to The Hacker News, the company counted about 16,000 extraction attempts from more than 4,000 accounts during a two-day spike, and tied a core cluster of that activity to people associated with Moonshot. OpenAI did not publish technical evidence, and that coverage carried no response from Moonshot.

Anthropic, Google and OpenAI have all argued that distillation lets competitors reproduce costly capabilities quickly and cheaply. Reuters notes that Anthropic said earlier this month that new defences had reduced the data harvested through such attacks by roughly 90 to 99 percent.

Why it matters

Distillation sits in a grey zone. Learning from another model's outputs is a common research technique, but AI providers object to it being used to build competing products, and Reuters describes it as a flashpoint in technology competition between the United States and China. If the report holds up, it would raise questions about how far recent Chinese gains depend on access to American frontier models. For now, the key facts remain unconfirmed, and none of the three companies has addressed them publicly.