DeepSeek, Moonshot trained new models on OpenAI’s GPT-6 Astra, report says

Key Points

1. The Information reported Chinese developers DeepSeek and Moonshot AI leaned on OpenAI's GPT-6 Astra outputs to train new models.
2. The technique, distillation, lets rivals reproduce advanced capabilities faster and far more cheaply, leading AI firms argue.
3. Distillation has become a flashpoint as US-China technology competition hardens and labs move to block data harvesting.

The latest:

Two of China’s most prominent AI developers relied heavily on outputs from OpenAI’s GPT-6 Astra model to train their newest systems, The Information reported Wednesday, citing people with direct knowledge. The outlet named Moonshot AI’s Kimi K3.1 Pro, released last week, and DeepSeek’s upgraded V4.1-Flash 1017, launched October 17. OpenAI, DeepSeek and Moonshot did not immediately respond to Reuters requests for comment.

Details:

• The technique: The practice is known as distillation, in which one model is trained on the outputs of another, more capable system. Leading AI firms argue it allows competitors to reproduce advanced capabilities far more quickly and at lower cost than building them from scratch.
• The models named: According to The Information, Moonshot AI’s Kimi K3.1 Pro, released last week, and DeepSeek’s upgraded V4.1-Flash 1017, launched on October 17, both benefited from distilling GPT-6 Astra. The report cited people with direct knowledge of the matter.
• OpenAI’s position: In a blog post dated September 30, OpenAI said it had linked a core part of a distillation campaign that began in July to individuals associated with Moonshot AI. The company did not name those individuals or describe what action it took against them.
• No response: OpenAI, DeepSeek and Moonshot AI did not immediately respond to Reuters requests for comment. None of the three has publicly addressed the specific claim that Astra outputs were used to train the two newly released Chinese models.
• Industry front: Anthropic, OpenAI and Google have all argued that distillation lets rivals copy advanced capabilities more quickly and cheaply. The complaint has shifted from a technical dispute among labs into a commercial and political argument over who captures the returns on frontier research.
• The countermeasures: Earlier this month, Anthropic said new measures designed to stop rivals from copying the capabilities of its Claude models had cut the volume of data harvested in so-called distillation attacks by roughly 90% to 99%. Anthropic did not detail how the measures work.
• The geopolitics: The practice has become a flashpoint as US-China technology competition intensifies. Washington’s export controls target the chips needed to train frontier models, but distillation operates on the output side, where hardware restrictions have no direct effect.
• The sourcing: The reporting rests on The Information’s account and OpenAI’s September blog post. DeepSeek and Moonshot have not conceded the use of Astra outputs, and the claim stands as a report attributed to people with direct knowledge.

Background:

DeepSeek drew global attention by releasing competitive models at a fraction of the cost of US rivals, a gap Western labs have repeatedly attributed in part to distillation. Moonshot AI, the Beijing-based developer behind the Kimi series, is the second Chinese lab named in OpenAI’s public statements on the practice.

Between the lines:

OpenAI’s September post already tied a distillation campaign to people associated with Moonshot, so Wednesday’s report extends an accusation rather than opening one. Anthropic’s claim that its countermeasures cut harvested data by up to 99% suggests leading labs now treat distillation as an engineering problem to be blocked at the interface, not only a grievance to be litigated after the fact.

What’s next

Watch for responses from OpenAI, DeepSeek or Moonshot AI, for further OpenAI disclosures of the kind published September 30, and for whether other US labs follow Anthropic in announcing technical countermeasures against distillation.