Report: DeepSeek, Moonshot AI trained new models on GPT-6 Astra output

DeepSeek and Moonshot AI reportedly used output from OpenAI's GPT-6 Astra as training data for their latest models, The Information reported on October 28, 2026. The Chinese labs had previously gathered outputs from Anthropic's Claude models intensively, according to the report.

Moonshot AI drew on Astra's outputs while developing Kimi K3.1 Pro, released the previous week, according to three people familiar with the matter cited by The Information. DeepSeek used Astra data to improve V4.1-Flash-1017, released on October 17, according to two other sources.

The technique involved is distillation, in which the responses or reasoning outputs of a stronger model are used as training data. The scale and method of Astra data use in the two companies' latest models have not been disclosed.

Anthropic said in an official blog post on October 12 that a new defense technique cut the data available to distillation attackers by 90-99% in its own evaluation. Earlier, on September 30, OpenAI announced it had detected an organized model-distillation attack involving people associated with Moonshot AI, saying attackers tried to extract models' private reasoning information using large numbers of accounts starting in July 2026 and that it blocked the activity. That incident occurred before Astra's release.

Moonshot AI declined to comment on the Astra distillation allegations, and DeepSeek did not respond to requests for comment, according to The Information.