Chinese laboratories, specifically DeepSeek and Moonshot AI, utilized responses from GPT-6 Astra, an advanced model developed by OpenAI, to train their latest versions of artificial intelligence, as reported by The Information.
This methodology, known as distillation, serves to replicate the functionalities of AI models using less computational power. However, the use of proprietary systems by competitors has drawn criticism from American companies, especially in the context of technological rivalry between the United States and China.
Sources with direct knowledge of the matter informed The Information that Moonshot AI used Astra responses during the development of Kimi K3.1 Pro, launched the previous week. Similarly, V4.1-Flash 1017, an update to DeepSeek's V4.1-Flash model, also benefited from this technique.
Distillation involves using one model's responses to instruct another. Both OpenAI and Anthropic apply this method when creating smaller, more agile, and cost-effective versions of their own systems. The controversy arises when external entities attempt to replicate the capabilities of proprietary models without permission.
To achieve this, these laboratories may try to extract the internal reasoning steps of the systems, information that developers do not intentionally disclose. According to sources consulted by the publication, collecting valuable training data can require more than ten times the volume of requests usually used, while also necessitating the use of multiple accounts to bypass detection mechanisms.
Distillation is not the only factor responsible for the progress of Chinese models. Restrictions imposed by the US on the export of sophisticated chips and resource limitations have also driven companies like DeepSeek to seek ways to develop and operate systems with lower computational capacity.
Until recently, Anthropic's Claude models were the main focus of this type of activity by Chinese laboratories, according to The Information. The publication points out that the shift in focus to Astra is partly due to the strengthening of defenses implemented by Anthropic in recent months.
In early October, Anthropic announced that new measures reduced collected data in distillation attempts by between 90% and 99%. These protections make it difficult to extract Claude's internal reasoning through intermediary networks, whose use violates the company's guidelines. Furthermore, repeated attempts can lead to stricter access restrictions and bans.
Anthropic detailed in documents related to its future Initial Public Offering (IPO) that clients who violated its terms represented approximately 5% to 7% of revenue in the first nine months of 2026. An intermediary network operator in China calculated that clients of this specific service contributed about $10 million monthly to the company's revenue.
At the end of September, OpenAI declared that it had detected and terminated a coordinated distillation campaign that began in July and was linked to individuals associated with Moonshot AI. The company alleged that those involved attempted to access the protected reasoning of its models through interactions that made this content available.
An OpenAI spokesperson assured The Information that the company treats the issue with extreme seriousness and is continuously improving its defenses. The company did not provide an estimate of the impact of this distillation on its revenue.
Moonshot did not issue a statement regarding the report, and DeepSeek did not respond to the request for comment. Both corporations did not publicly respond to the accusations mentioned in the publication.
The Information also reported that Liang Wenfeng, CEO of DeepSeek, initially opposed resorting to distillation in the development of version V4. However, according to two sources familiar with the process, he eventually yielded to competitive pressure after the advancement of GLM models from competitor Z.ai, which were also reportedly created using this technique.