Chinese developers DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's GPT-6 Astra to train their latest models, The Information reported on Wednesday, according to Reuters, citing people with direct knowledge of the matter. The models named are Moonshot's Kimi K3.1 Pro, released last week, and DeepSeek's upgraded V4.1-Flash 1017, launched on October 17. None of the three companies immediately responded to Reuters requests for comment, so for now this is a well-sourced report, not an admitted fact, and it should be read that way.
The technique at issue is distillation: training a smaller or cheaper model on the outputs of a stronger one, so the student inherits much of the teacher's reasoning style and capability at a fraction of the original training cost. Distillation itself is a legitimate, everyday tool when you own the teacher or have a licence to use its outputs. The fight starts when the teacher is a rival's frontier model, accessed through ordinary API accounts at scale. OpenAI, Anthropic and Google have all argued that this lets competitors reproduce advanced capabilities far more quickly and cheaply than building them from scratch.
This report does not arrive in a vacuum. In a September 30 blog post, OpenAI said it had linked a core part of a distillation campaign that began in July to individuals associated with Moonshot AI. Earlier this month, Anthropic said new measures designed to stop rivals copying Claude's capabilities had cut the data harvested in so-called distillation attacks by about 90 to 99 percent. In other words, the major labs are no longer treating output harvesting as a terms-of-service footnote. They are instrumenting their APIs to detect it, throttling it, and naming names in public.
The economics explain the pressure. The same week, Microsoft disclosed in a regulatory filing, reported by TechCrunch, that it recorded 10 billion dollars in revenue from its commercial arrangements with OpenAI in the quarter ended September 30, nearly 11 percent of Microsoft's total quarterly revenue of 92.3 billion dollars, with 7.2 billion dollars still owed to it by OpenAI at quarter end. Frontier outputs are worth that kind of money because frontier capability is expensive to create. If a rival can siphon a meaningful slice of that capability through API calls priced for ordinary customers, the return on building the frontier model in the first place starts to leak away.
For people building with AI, the practical lessons are immediate. Treat model outputs as a licensed asset, not a free raw material: check the terms of every provider you use before training on its outputs, and keep records of where your training data came from. Expect providers to get better and more aggressive at detecting large-scale extraction, including for customers who are doing it innocently through aggressive caching or synthetic-data pipelines that look like harvesting. And if your own product exposes a capable model through an API, assume someone is already trying to distil it, and build the rate limits, logging and output controls you would want in place before, not after, a report like this one names your model as the teacher.
