DeepSeek and Moonshot relied on OpenAI outputs to train new models, report says

DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra to train their latest models, The Information reported, citing people with direct knowledge. Reuters relayed the allegations on October 28, 2026, putting two recent releases at the center of a dispute over rivals using advanced AI systems to improve their own.

The models named were Moonshot’s Kimi K3.1 Pro and DeepSeek’s upgraded V4.1-Flash 1017. Kimi K3.1 Pro had been released the previous week; DeepSeek’s update launched October 17. The Information said both benefited from distilling Astra—training another model using responses generated by OpenAI’s system.

OpenAI, DeepSeek and Moonshot did not immediately respond to Reuters’ requests for comment. The Decoder also covered the allegations on October 29, citing The Information.

Answers and hidden reasoning are different targets

OpenAI, Anthropic and Google have argued that the practice lets competitors reproduce advanced capabilities faster and at lower cost.

OpenAI has described a more specific concern: extracting protected reasoning, the internal record a model uses to work through a task. That record can contain information omitted from its final answer. In its September 30 security disclosure, OpenAI called systematic, unauthorized use of outputs or reasoning to train another model “adversarial distillation.”

The earlier Moonshot link has limits

That earlier disclosure attributed a core cluster of extraction activity to individuals associated with Moonshot AI. OpenAI explicitly said it was unclear whether all operators came from a single actor.

One technique involved moving encrypted reasoning from one conversation into another and asking a model to decrypt and transcribe it. OpenAI said the operators manipulated model interactions to expose protected material. Its account described an extraction route, not evidence that those July attempts trained either model named in the new report.

The operators did not break our encryption, compromise a database, or gain direct access to stored user conversations.

OpenAI, September 30 security disclosure

Defending the reasoning, not just the accounts

The company’s stated concern extends beyond competitive copying. OpenAI warned that extracted reasoning could train another model without preserving safeguards attached to the original system’s visible answers. That is a potential safety consequence, not a demonstrated outcome for Kimi K3.1 Pro or V4.1-Flash 1017.