The Chinese laboratories DeepSeek and Moonshot AI turned to responses from GPT-6 Astra, OpenAI’s advanced model, to train their latest versions, according to a report by The Information.
The practice, known as distillation, helps reproduce the capabilities of artificial intelligence (AI) models with fewer computational resources. However, competitors’ use of proprietary systems has drawn criticism from American companies amid the technological rivalry between the United States and China.
DeepSeek’s Kimi K3.1 Pro and V4.1-Flash 1017 models were identified as beneficiaries of GPT-6 Astra data in a recent report. – Image: mundissima/Shutterstock
Chinese models turned to GPT-6 Astra
According to people with direct knowledge of the matter heard by The Information, Moonshot AI used Astra responses during the development of Kimi K3.1 Pro, launched last week. The V4.1-Flash 1017, an update to DeepSeek’s V4.1-Flash model released on October 17, also benefited from the technique.
Distillation consists of using the responses of one model to train another. OpenAI and Anthropic apply the method when developing smaller, faster, and cheaper versions of their own systems.The controversy arises when outside companies attempt to replicate the capabilities of proprietary models without authorization.
To do so, labs may try to extract internal reasoning steps from systems—information the developers deliberately do not disclose. According to sources interviewed by the publication, collecting data useful for training may require more than ten times the number of requests normally used, as well as multiple accounts to evade detection mechanisms.
Distillation alone does not explain the advancement of Chinese models. U.S. restrictions on exports of advanced chips and resource constraints also prompted companies such as DeepSeek to seek ways to develop and operate systems with less computing capacity.
Anthropic strengthened protections against distillation
Until recently, Anthropic’s Claude models were the primary targets for this type of activity among Chinese labs, according to The Information. The publication attributed the shift toward Astra partly to Anthropic’s strengthening of its defenses in recent months.
In early October, the company said new measures had reduced the amount of data collected through distillation attacks by between 90% and 99%. The protections make it more difficult for networks of intermediaries to extract Claude’s internal reasoning, whose use violates the company’s policies. Repeated attempts may also result in stricter access limits and bans.
Anthropic said in documents related to its planned initial public offering of shares that customers who violated its terms accounted for approximately 5% to 7% of revenue in the first nine months of 2026.A Chinese proxy network operator estimated that customers of this type of service contributed approximately US$10 million a month to the company's revenue.
OpenAI said it had identified a coordinated distillation campaign launched in July and linked to individuals associated with Moonshot AI. – Image: gguy/Shutterstock
OpenAI linked campaign to people connected to Moonshot
At the end of September, OpenAI said it had identified and stopped a coordinated distillation campaign launched in July, which it linked to individuals associated with Moonshot AI. According to the company, those involved attempted to extract the protected reasoning from its models through interactions that made that content accessible.
An OpenAI spokesperson told The Information that the company takes the issue very seriously and continues to improve its defenses. The company did not estimate the impact of distillation on its revenue.
Moonshot did not comment on the report, and DeepSeek did not respond to a request for comment. Neither company publicly responded to the allegations cited in the publication.
The The Information also reported that DeepSeek CEO Liang Wenfeng initially resisted resorting to distillation in developing the V4 generation. According to two sources with knowledge of the process, he eventually gave in to competitive pressure after Chinese competitor Z.ai advanced its GLM models, which may also have been developed using the technique.
Valdir Antonelli is a journalist specializing in digital marketing and consumption.
Ana Figueiredo is Olhar Digital's technology editor. She has a degree in Journalism from the Federal University of Uberlândia (UFU).
