Two leading Chinese artificial intelligence developers, DeepSeek and Moonshot AI, relied heavily on answers from OpenAI's GPT-6 Astra to train their newest models, The Information reported Wednesday, citing people with direct knowledge, according to Reuters. OpenAI, DeepSeek and Moonshot did not respond to Reuters' requests for comment, and none of them has confirmed the account. If accurate, it would add to signs that rivals can copy much of a frontier model's ability through its paid interface, which bears on how far chip export controls alone can protect the U.S. lead.
What Happened
The practice at issue is distillation: training one AI model on the answers another model produces, so the new model cheaply picks up the older one's skills. According to The Information's sources, as relayed by Reuters, the models that benefited were Moonshot's Kimi K3.1 Pro, released last week, and DeepSeek's V4.1-Flash 1017, launched Oct. 17.
In a Sept. 30 blog post, OpenAI said it had tied a core part of a distillation campaign that began in July to people associated with Moonshot. The Primary Wire reported that finding, citing CyberScoop, which noted that OpenAI had published no technical evidence for the attribution. Earlier this month, Anthropic said new safeguards had cut the data harvested from its Claude models in distillation attacks by roughly 90% to 99%, as The Primary Wire reported, citing Reuters.
Why It Matters
Distillation has become a point of friction between U.S. and Chinese AI developers. Anthropic, OpenAI and Google argue that it lets rivals reproduce advanced capabilities faster and more cheaply than building them, according to Reuters.
For policymakers, the question is what export controls can achieve. Washington has limited sales of Nvidia's most advanced chips to China since 2022, and U.S. officials are examining distillation as one way Chinese labs may keep pace despite those limits, The Primary Wire reported in September, citing PhoneWorld. If a model's skills can be harvested through ordinary customer accounts, protecting them depends on identity checks, abuse detection and contract enforcement as much as on hardware.
For companies, the stakes run both ways. Businesses that build on Chinese open models, which handled a majority of tokens on the OpenRouter and Vercel developer gateways in data cited by that same September story, may face harder questions about how those models were made. Federal agencies already have to drop DeepSeek: the Department of Homeland Security has ordered civilian agencies to remove its products, including third-party tools built on its technology, The Primary Wire reported Oct. 20.
What to Watch
Watch whether OpenAI confirms the report, closes accounts or names those responsible, as it did with Moonshot-linked users last month, and whether DeepSeek or Moonshot responds. Also watch whether OpenAI tightens access to its models' outputs; researchers recently found a way around one of its earlier fixes, The Primary Wire reported, citing SecurityWeek. In Washington, the signal is whether the Commerce Department or Congress takes up distillation directly, through export rules, restrictions on specific firms or legislation.
