Chinese artificial intelligence developers DeepSeek and Moonshot AI actively used outputs from OpenAI's GPT-6 Astra model to train their newest systems. This was reported on October 31 by The Information and Reuters, citing sources familiar with the internal processes of both companies. This concerns the practice of distillation—a method where responses from a more powerful model are used as training material for another system, allowing it to replicate certain AI capabilities without comparable computing costs.
The distillation method and its economic logic
Distillation in the context of generative AI involves sending a large number of prompts to a stronger model, collecting its responses, and subsequently training a less resource-intensive system on that data. According to sources at The Information, this exact approach allowed DeepSeek and Moonshot AI to accelerate the development of new models without investing amounts comparable to OpenAI's expenses for creating GPT-6 Astra. For companies whose computing power lags behind the market leader, this method serves as a tool to narrow the capability gap without the need to train models from scratch on their own.
Specific models and timeline
Reuters clarifies that GPT-6 Astra distillation results were used in training Moonshot AI's Kimi K3.1 Pro model, released last week, as well as the updated DeepSeek V4.1-Flash 1017 version, which appeared on October 17, 2026. Thus, both Chinese companies applied the same method within updates released over the past two weeks, indicating a systematic rather than isolated use of GPT-6 Astra data.
History of the data collection campaign and OpenAI's position
The use of distillation has previously been the subject of public disputes. In a publication dated September 30, 2026, OpenAI reported that it discovered participants in a data collection campaign starting in July, which the company claimed was linked to Moonshot AI. According to ixbt.com citing the same material, OpenAI recorded activity from over 15,000 users, blocked the involved accounts by July 28, halted the campaign, and transferred the collected data to other AI model developers. OpenAI, Anthropic, and Google consistently oppose the practice of distillation, believing it allows competitors to copy their model capabilities without financial investments in research and development.
Protection measures and lack of comments
Earlier in October 2026, Anthropic stated that new defense measures for its Claude models reduced the volume of data collected during distillation attacks by 90–99%. This indicates that leading US labs view distillation as a real threat vector to intellectual property and are actively investing in technical countermeasures. Representatives of OpenAI, DeepSeek, and Moonshot AI did not respond to Reuters requests for comment on the October 31 publications by The Information and Reuters. Neither of the Chinese companies denied or confirmed the use of GPT-6 Astra data in training their models.
