DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra to train two models released in October, Techzine reported on 29 October. The report names Moonshot AI’s Kimi K3.1 Pro and DeepSeek’s V4.1-Flash 1017, placing a familiar model-training technique at the centre of a dispute over access to commercial AI systems.
Key points
- The Information reported that both companies drew heavily on GPT-6 Astra responses when training their latest models.
- Moonshot AI released Kimi K3.1 Pro in the week before 28 October; DeepSeek launched V4.1-Flash 1017 on 17 October.
- OpenAI has linked an earlier distillation campaign to individuals associated with Moonshot AI, while Anthropic says its defences have sharply reduced extraction.
GPT-6 Astra and the two October releases
Moonshot AI released Kimi K3.1 Pro in the week before 28 October, and DeepSeek launched V4.1-Flash 1017 on 17 October, Techzine reported. The Information attributed its account of their training to people with direct knowledge of the matter. Its central allegation concerns the source of some training material: responses generated by OpenAI’s model, which the two companies are reported to have used extensively.
The method is called distillation. A model receives examples of another model’s responses and learns to produce answers of its own from them, much as a pupil might practise with worked examples rather than the book used to write them. The technique has ordinary uses in model development, including teaching a smaller system from a more capable one. Using a commercial service to gather responses at scale for that purpose may breach the provider’s terms, although the training method itself does not establish whether any particular use was authorised.
Drafting an answer to a routine question could draw on a model trained with another model’s responses. If that training worked as intended, the answer could reflect capabilities learned from those examples even though the finished text would come from the model being used.
For companies building AI products, the distinction between obtaining an answer and collecting answers to train a competing system matters. Both activities can involve sending requests to a model, but the second turns its replies into material for another model’s development. The reported use of GPT-6 Astra concerns that second activity. Techzine said large-scale distillation through a commercial API may violate its terms of service.
OpenAI’s July campaign account
OpenAI had raised a more specific concern before the October releases. In a 30 September blog post, it said it traced a central element of a July distillation campaign to individuals associated with Moonshot AI. AI Affairs reported OpenAI’s link to Moonshot AI individuals. The company’s account concerned a campaign it had investigated; The Information’s later report concerned the training of Kimi K3.1 Pro and V4.1-Flash 1017.
OpenAI said the activity began on 1 July, Times of AI reported. The company counted 16,000 requests matching a relevant extraction pattern from more than 4,000 users. Its investigation also identified related activity across more than 15,000 users, which OpenAI said it had fully disrupted by 28 July.
Those counts describe requests and accounts, rather than a measured amount of training data obtained: OpenAI said an attempted extraction was not necessarily successful. According to its account, the campaign tried to expose protected reasoning by manipulating interactions with the model. It said the activity did not involve breaking encryption or obtaining direct access to stored conversations. OpenAI responded by restricting accounts, tightening technical controls and expanding its monitoring of coordinated activity.
Anthropic’s October distillation defences
Anthropic described its own measures against distillation on 12 October. They included detecting suspicious requests and protecting reasoning transcripts, the internal steps a model can produce while working towards an answer. AI Affairs covered Anthropic’s account. Restricting access to those transcripts would limit one kind of material that could be collected to teach another model.
Anthropic said its guardrails had reduced data taken in distillation attacks by about 90 to 99 percent and that it had permanently blocked tens of thousands of malicious accounts, Techzine reported. Times of AI gave the company’s reduction figure as around 92-99%. Anthropic also acknowledged that its measures could sometimes affect legitimate users, according to Times of AI.
OpenAI, DeepSeek and Moonshot AI did not immediately respond to Reuters’ requests for comment on the reported GPT-6 Astra training.
