DeepSeek and Moonshot AI used outputs from OpenAI’s GPT-6 Astra model to train their latest models, The Information reported on Wednesday (Oct. 28), citing people with direct knowledge of the matter, according to Reuters. The Information’s article, which is behind a paywall, carries a headline describing Astra as the top distillation target for China’s AI labs.
Reuters said The Information reported that Moonshot’s Kimi K3.1 Pro model, released last week, and DeepSeek’s upgraded V4.1-Flash 1017 model, launched on October 17, benefited from distilling Astra. Distillation is the practice of using one model’s outputs to help train another. Reuters described it as a flashpoint for leading AI companies as U.S.-China technology competition intensifies. Reuters said firms argue it lets rivals reproduce advanced capabilities more quickly and cheaply.
OpenAI, DeepSeek and Moonshot AI did not immediately respond to Reuters requests for comment, Reuters said. Every claim about Astra in this account traces to The Information’s unnamed sources, relayed by Reuters, and none of the three companies had commented when Reuters published. Reuters’ summary described the models as relying heavily on Astra outputs; The Information’s article is paywalled, so how it characterizes the extent of the use could not be checked here.
Reuters noted a separate OpenAI statement. 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. In that post, OpenAI described the activity as an effort to extract protected reasoning from its models. It said operators did not break its encryption or reach stored user conversations, and that it was unclear whether all operators came from a single actor. It said it fully disrupted a related cluster of more than 15,000 users by July 28. OpenAI’s post does not describe the use of Astra outputs to train Kimi K3.1 Pro or V4.1-Flash 1017, which is the subject of The Information’s report. Details of the September campaign are in an earlier report.
Moonshot said on Saturday, October 24, according to a Reuters-sourced report, that it had reduced the cyber-offensive capabilities of Kimi K3.1 Pro and planned to publish its weights by November 4. That report did not address distillation.
Reuters added that Anthropic said earlier this month that new measures designed to prevent rivals from copying Claude’s capabilities have reduced data harvested in distillation attacks by about 90% to 99%. Anthropic’s Oct. 12 post said the number of reasoning tokens extracted by distillers is consistently 10 to 100 times lower than before. Background on the term is in a distillation explainer.
Sources
- Reuters via The Hindu BusinessLine: DeepSeek, Moonshot used OpenAI’s GPT-6 Astra to train latest models, Report (October 29, 2026; Reuters dateline Oct. 28)
- The Information: OpenAI’s Astra Model Becomes Top Distillation Target for China’s AI Labs (paywalled; headline and byline only)
- OpenAI: Disrupting a coordinated model-distillation campaign (September 30, 2026)
- Anthropic: Strengthening our defenses against distillation attacks (October 12, 2026)