Chinese artificial intelligence companies DeepSeek and Moonshot AI reportedly relied heavily on outputs from OpenAI’s GPT-6 Astra to help train their latest AI models, according to a report by The Information published on October 28, 2026.
The report, citing people with direct knowledge of the matter, highlights growing competition among leading artificial intelligence developers in the United States and China. It also brings renewed attention to a controversial training technique known as AI model distillation, which can help developers reproduce certain capabilities of more advanced models at a lower cost.
The reported use of GPT-6 Astra raises important questions about intellectual property, AI development costs, access to advanced technology, and the measures companies are taking to protect their models.
However, the allegations have not been independently established as a legal violation, and the companies involved had not immediately responded to Reuters’ requests for comment when its report was published.
Which AI Models Were Reportedly Involved?
According to The Information, two recently released AI models benefited from training using outputs from OpenAI’s GPT-6 Astra.
The first is Moonshot AI’s Kimi K3.1 Pro, which was released during the week before the report. The second is DeepSeek’s V4.1-Flash 1017, an upgraded version launched on October 17, 2026.
People familiar with the model-development efforts reportedly told The Information that both companies used outputs from GPT-6 Astra as part of their training processes.
The report did not publicly establish exactly how many AI-generated responses were used, how the data was collected, or how much the technique contributed to the performance of either model.
This distinction matters because AI systems are developed through multiple processes, including training on datasets, model architecture improvements, optimization techniques, and extensive testing.
Using another model’s outputs may contribute to development, but it does not necessarily explain a model’s entire performance or capabilities.
Source: The Information — OpenAI’s Astra Model Becomes Top Distillation Target for China’s AI Labs.
What Is AI Model Distillation?
AI model distillation is a technique in which a model learns from the outputs of another model, often one that is more capable or more expensive to operate.
In a typical distillation process, a developer sends questions or tasks to a more advanced model and collects its responses. Those responses can then become training examples for another model, helping it learn patterns in how the more capable system solves problems or generates answers.
For example, a developer could ask a powerful AI model to explain scientific concepts, solve mathematics problems, summarize documents, or generate computer code. The resulting answers could be used as examples to train a different model.
This approach can help developers improve a model without having to reproduce every aspect of the original model’s training process.
Distillation can also help create smaller systems that are faster or cheaper to run while retaining useful capabilities.
However, the quality of the resulting model depends on many factors, including the training data, the learning method, the model’s design, and the quality of its evaluation.
Why Is Distillation Controversial?
Distillation is not inherently improper. AI companies, including major developers, use forms of distillation to create smaller or more efficient versions of their own systems.
The controversy arises when an organization attempts to extract capabilities from a proprietary model without authorization or in violation of the model provider’s terms.
Companies that invest heavily in developing advanced AI systems argue that large-scale extraction of their models’ outputs can allow competitors to benefit from their research and development efforts without bearing comparable costs.
OpenAI and Anthropic have both raised concerns about attempts to extract the capabilities of their models through coordinated or large-scale activity.
The key issue is therefore not simply whether distillation occurred, but how the data was obtained, what restrictions applied, and whether those restrictions were violated.
Why Would AI Companies Use a More Advanced Competitor’s Model?
Developing advanced AI systems requires substantial investment in computing infrastructure, research, engineering, training data, and testing.
Distillation can provide a way to learn from an already capable model’s responses rather than relying entirely on a separate, resource-intensive development process.
For companies operating in a competitive market, this can potentially reduce development time and improve performance on selected tasks.
The technique may be particularly attractive when developers are trying to build models that compete with established systems in coding, mathematics, reasoning, and other demanding applications.
However, distillation is not a complete substitute for developing a capable AI system. Companies still need engineering expertise, suitable computing resources, high-quality training data, and methods for evaluating reliability and safety.
The recent report also does not establish that distillation was the sole reason for the capabilities of DeepSeek’s or Moonshot AI’s latest models.
OpenAI Has Previously Raised Concerns About Model Distillation
The report comes amid broader concerns about efforts to extract capabilities from advanced AI systems.
On September 30, 2026, OpenAI said it had linked a significant part of a distillation campaign that began in July to individuals associated with Moonshot AI.
The company described attempts to extract protected reasoning and other capabilities from its models. Such activity can involve systematically collecting large numbers of responses rather than using a model for ordinary individual tasks.
The latest reporting adds to those concerns by identifying GPT-6 Astra as a model reportedly used during the development of newer systems from DeepSeek and Moonshot AI.
Nevertheless, the earlier allegations and the latest report should not be treated as proof that every model developed by either company was trained improperly. The details of the specific activities, the methods used, and their contribution to the models’ performance remain important distinctions.
For background, readers can consult OpenAI’s official website for its published statements about model security and distillation.
How Does This Affect Competition Between the United States and China?
The development reflects the increasingly intense competition between American and Chinese AI companies.
Leading US developers have invested heavily in advanced models and supporting infrastructure. Chinese companies, meanwhile, have developed competitive AI systems and explored different methods to improve training efficiency and reduce operating costs.
Advanced semiconductor access, computing resources, research talent, and software optimization all influence the ability of companies to build and deploy capable models.
Distillation has become part of this wider debate because it may help developers reproduce certain capabilities more efficiently. At the same time, AI companies are seeking stronger protections against unauthorized extraction of their technology.
The dispute illustrates a difficult balance: developers want to encourage research and competition, but they also want to protect intellectual property and prevent the misuse of their services.
It is important, however, not to attribute the progress of Chinese AI developers to distillation alone. Research innovation, engineering improvements, model architecture, and efficient use of computing resources can also contribute to competitive performance.
What Could This Mean for the Future of AI Development?
The growing attention to model distillation could lead AI companies to introduce stronger safeguards around access to their systems.
Possible measures include monitoring unusual usage patterns, limiting automated extraction, detecting attempts to obtain restricted reasoning, and improving systems for identifying coordinated misuse.
Companies may also need to distinguish legitimate customers who use AI tools for research or productivity from users attempting to collect large volumes of outputs for model replication.
These protections could affect how developers access advanced AI services and how third-party organizations build products using them.
At the same time, distillation will likely remain an important technique for legitimate AI development, particularly when companies want to create smaller, more efficient models from systems they own or are authorized to use.
The challenge for the industry is to establish clear rules that protect proprietary technology without unnecessarily restricting legitimate research, innovation, and competition.
Conclusion
The Information’s report that DeepSeek and Moonshot AI used outputs from OpenAI’s GPT-6 Astra to train their latest models highlights a growing conflict over the use of advanced AI systems.
The technique, known as model distillation, can help developers reproduce useful capabilities more quickly and economically. It is also a legitimate training method in appropriate circumstances, including when developers distill their own models or have permission to use another system’s outputs.
The controversy centers on whether proprietary model capabilities were extracted in ways that violated applicable restrictions.
Although the report identifies models associated with DeepSeek and Moonshot AI, important details about the training processes remain undisclosed, and the claims should not be treated as a final legal finding.
As AI competition intensifies, the ability to build powerful models will remain important—but protecting intellectual property, maintaining transparency, and establishing responsible rules for model development will be increasingly significant as well.