DeepSeek and Moonshot AI reportedly relied heavily on answers generated by OpenAI's GPT-6 Astra to improve their latest models. The method is called distillation: one AI system produces examples that another learns from.

The Information reported the use of Astra on October 28, citing people with direct knowledge of the work. Its account named Moonshot's Kimi K3.1 Pro, released last week, and DeepSeek's V4.1-Flash 1017, launched on October 17.

Reuters' report said OpenAI, DeepSeek and Moonshot had not immediately responded to its requests for comment. Reuters did not give dataset sizes or say how much of each model's improvement came from Astra.

Learning from another model's answers

The basic process is straightforward. A developer sends prompts to a capable model, collects the answers and uses those examples to train another system. The transfer happens through generated responses, without needing the teacher model's underlying numerical settings, known as weights.

OpenAI described the approach in an October 2024 explanation:

"Model distillation involves fine-tuning smaller, cost-efficient models using outputs from more capable models, allowing them to match the performance of advanced models on specific tasks at a much lower cost."

The phrase "specific tasks" matters. Training on good examples can improve a model's performance on the work represented in those examples. It does not establish that the student has inherited every capability of the system providing the answers.

That workflow also needs testing. The same explanation describes evaluating the trained model, checking whether it meets the required standard and improving the dataset or training process when results fall short. A collection of impressive answers is a starting point for training, rather than a finished product.

The dispute is about permission and competition

OpenAI has offered distillation within its own platform. Its current services agreement also restricts customers from using outputs to develop AI models that compete with its products, subject to defined exceptions.

Those exceptions include certain internally used classification or data-organization models and fine-tuning or customization through OpenAI's own services. The agreement gives customers ownership of outputs to the extent allowed by law, while keeping these limits on their use.

There is a separate security concern when developers try to obtain reasoning that the original provider has chosen to keep hidden. In a September 30 statement, OpenAI described a coordinated extraction campaign first observed in July. On attribution, it said:

"It is unclear whether all operators we observed during the relevant time period originated from a single actor. However, we attribute a core cluster of the activity to individuals associated with Moonshot AI, the developer of Kimi."

That statement concerned the earlier campaign, not a published finding about the two new model versions. The company said operators manipulated model interactions to expose protected reasoning, and that they had not broken its encryption or obtained direct access to stored user conversations.

OpenAI said it responded with account restrictions, stronger protections around hidden reasoning and coordination with third-party providers when related activity passed through their services.

The defenses have costs too

Anthropic described its own response on October 12. It said early evidence suggested new safeguards had reduced data harvested by distillers by roughly 90%-99%. Its more detailed assessment focused on reasoning extracted from Claude on its own platform, rather than all copying across the industry.

The company said it had strengthened controls around reasoning, added systems to detect extraction attempts and banned associated accounts. It also cautioned that it did not know how long the gains would last.

Those measures affect ordinary developers as well. Some customers had to update their integrations, legitimate requests could occasionally be blocked, and certain blocked requests now incur charges. Protecting model capabilities can therefore change the cost and experience of using the same tools that paying customers depend on.