Chinese artificial intelligence developers DeepSeek and Moonshot AI reportedly relied on outputs from OpenAI’s GPT-6 Astra to improve their latest models, according to a report by The Information published on October 28, 2026. The report says Moonshot used Astra outputs while developing Kimi K3.1 Pro, while DeepSeek’s V4.1-Flash 1017 upgrade also benefited from the technique. The allegations arrive amid growing competition between US and Chinese AI companies over model capabilities, computing resources and access to advanced AI systems.
The reported activity involves model distillation, a technique that allows developers to train a model using responses generated by another, often more capable model. Although distillation is a standard AI development method, using proprietary models to extract capabilities without authorization can violate service terms and raise concerns about intellectual property. OpenAI has previously reported attempts to extract protected reasoning from its models, including a campaign it linked to individuals associated with Moonshot AI. However, the latest reporting does not establish precisely how much Astra contributed to either Chinese model or whether every reported activity violated applicable terms.
Key Takeaways
- Reported use of Astra: DeepSeek and Moonshot AI reportedly used GPT-6 Astra outputs to help train their latest AI models.
- Models involved: The report identifies Moonshot’s Kimi K3.1 Pro and DeepSeek’s V4.1-Flash 1017.
- Distillation explained: Developers use outputs from a stronger model to improve another model’s capabilities, potentially reducing training costs and development time.
- OpenAI’s response: The company previously reported disrupting a distillation campaign linked to individuals associated with Moonshot AI.
- Industry-wide issue: Anthropic and other AI developers have also raised concerns about unauthorized extraction of proprietary model capabilities.
- Unresolved questions: The extent of Astra’s contribution, the precise methods used and the responses from the companies involved remain unclear.
What Does the Report Say About DeepSeek and Moonshot AI?
According to The Information, both Chinese AI developers used outputs from OpenAI’s GPT-6 Astra during the development of recent models.
Moonshot AI reportedly used Astra during work on Kimi K3.1 Pro, released the week before the report. DeepSeek’s V4.1-Flash 1017, an October 17 upgrade to its V4.1-Flash model, also reportedly benefited from distillation.
The report cites people with direct knowledge of the activity. Reuters subsequently reported the allegations, noting that OpenAI, DeepSeek and Moonshot AI had not immediately responded to its requests for comment.
The claims are significant because Astra is described in the report as one of OpenAI’s most capable models. Access to its outputs could help another developer study how a leading model handles complex reasoning, coding and other tasks.
However, the report does not establish that either company copied Astra’s model weights or obtained direct access to OpenAI’s internal systems. Distillation typically relies on a model’s outputs rather than access to its underlying parameters.
That distinction matters: learning from a model’s responses is technically different from stealing its source code, model weights or training database.
What Is AI Model Distillation?
Model distillation is a training technique in which a smaller or different model learns from outputs produced by another model, often called the teacher model.
The process can help developers reproduce useful capabilities without repeating the entire process of building a frontier model from scratch.
A simplified example illustrates the approach:
| Stage | How distillation works |
|---|---|
| 1. Generate questions | A developer prepares prompts covering reasoning, coding or other capabilities |
| 2. Obtain responses | A teacher model produces answers to those prompts |
| 3. Build training data | The answers are collected and prepared for training |
| 4. Train another model | The student model learns patterns from the generated examples |
| 5. Evaluate performance | Developers test the student model against benchmarks and real tasks |
Distillation can reduce the resources required to develop a capable model, although its effectiveness depends on the quality and diversity of the training data, the student model’s architecture and the training method.
The technique is not inherently improper. AI companies use distillation to create smaller, faster and less expensive versions of their own systems.
The controversy arises when developers extract capabilities from a competitor’s proprietary model through methods that violate access restrictions or contractual terms.
Why OpenAI Is Concerned About Distillation
OpenAI and other frontier AI companies invest heavily in model training, computing infrastructure, research and safety systems. Their proprietary models represent a substantial commercial investment.
If a competitor can use large volumes of generated responses to reproduce some of those capabilities, it may be able to reduce the time and resources needed to develop its own systems.
OpenAI has also expressed concern about attempts to extract protected reasoning—the internal processes that a model uses to work through a task but that are not intended to be exposed in ordinary responses.
On September 30, 2026, OpenAI published a statement describing a coordinated model-distillation campaign that began in July. The company said it had linked a core cluster of the activity to individuals associated with Moonshot AI and had taken steps to disrupt it.
OpenAI said the campaign involved attempts to elicit protected reasoning through model interactions. It also emphasized that the activity did not involve breaking encryption or directly compromising its stored user conversations.
The company’s position is that unauthorized extraction can transfer valuable capabilities without the original developer’s permission. OpenAI has said it is strengthening account controls, technical protections and industry information-sharing to detect such activity.
The September disclosure provides context for the latest report, but it should not be treated as independent confirmation of every detail about Astra’s use in the two newly identified models.
How Anthropic Is Responding to Similar Activity
OpenAI is not the only company confronting this issue. Anthropic has also reported large-scale attempts to extract capabilities from its Claude models.
In an earlier disclosure, Anthropic attributed campaigns to DeepSeek, Moonshot AI and MiniMax, saying they collectively generated more than 16 million exchanges with Claude through approximately 24,000 fraudulent accounts. The company said the activity targeted areas including reasoning, coding, data analysis and AI-agent capabilities.
Anthropic subsequently reported that new defenses had reduced the amount of data harvested in distillation attacks by approximately 90% to 99%.
These are Anthropic’s own findings and performance claims, rather than independently established measurements across every AI provider.
The reports nevertheless illustrate a shared industry problem. As frontier models become more capable, their outputs become valuable training resources. Providers must find ways to offer useful services while limiting attempts to systematically extract proprietary capabilities.
What This Means for China’s AI Industry
Chinese AI developers have gained international attention for producing competitive models despite constraints involving access to advanced chips and computing resources.
Distillation is one technique that can help developers make better use of available resources. High-quality model outputs can provide examples of complex problem-solving, coding and reasoning that are useful when training other systems.
However, distillation alone does not explain the progress of Chinese AI companies. Model architecture, training data, optimization, inference efficiency and access to computing infrastructure also influence performance.
There is also a distinction between producing a model that performs well on selected benchmarks and developing a reliable, general-purpose system. The quality of a model must be evaluated across a range of tasks, safety tests and real-world applications.
For DeepSeek and Moonshot AI, the reported use of Astra raises questions about the contribution of external model outputs to their latest releases. The available reporting does not quantify that contribution or establish how much of each model’s performance improvement came from distillation.
Could Distillation Change the AI Business Model?
The dispute highlights a tension in the AI industry. Model providers want developers to use their products, but they also want to prevent competitors from using those services as a substitute for expensive research and training.
If large-scale extraction becomes easier, companies may respond with stricter usage limits, better detection systems, stronger account verification and restrictions on suspicious patterns of activity.
Those protections could help defend proprietary capabilities, but they may also make it more difficult for legitimate researchers and developers to access advanced models. Providers will need to distinguish ordinary use, authorized research and internal distillation from coordinated attempts to extract capabilities in violation of their policies.
For businesses building AI products, the dispute also reinforces the importance of understanding API terms, data-use restrictions and model licensing conditions before incorporating third-party outputs into commercial training pipelines.
The Bigger Picture
The reported use of GPT-6 Astra by DeepSeek and Moonshot AI highlights how AI competition increasingly extends beyond building models from scratch. Access to the outputs of a strong system can provide a valuable source of training data, potentially helping developers improve their own products more quickly and at lower cost.
At the same time, the controversy shows why model providers are investing in safeguards against unauthorized extraction. The challenge is to protect proprietary capabilities without undermining legitimate research, interoperability and the beneficial use of distillation within the AI ecosystem.
FAQs
1. Did DeepSeek and Moonshot AI use OpenAI’s GPT-6 Astra?
A report by The Information, cited by Reuters on October 29, 2026, says both companies relied on Astra outputs while developing recent models. The claims are based on sources with direct knowledge of the matter and have not been independently detailed in public disclosures by the companies involved.
2. What is model distillation in AI?
Model distillation is a technique in which one model learns from examples or outputs produced by another model. It can help developers create models that reproduce useful capabilities more efficiently than training entirely from scratch.
3. Is AI model distillation illegal?
Not automatically. Distillation is a legitimate and widely used training technique. Whether a particular activity violates rules or laws depends on factors such as authorization, service terms, intellectual property rights and the methods used to obtain the training data.
4. Did DeepSeek or Moonshot steal OpenAI’s model weights?
The cited report describes the use of Astra’s outputs for training. It does not establish that either company stole OpenAI’s model weights, source code or internal training database.
5. Why are OpenAI and Anthropic trying to prevent distillation attacks?
They argue that unauthorized extraction can allow competitors to reproduce valuable capabilities without permission and potentially bypass the costs and safeguards associated with developing advanced models. Both companies have reported efforts to detect suspicious usage and restrict coordinated extraction campaigns.
Looking Ahead
The next important development will be whether OpenAI, DeepSeek or Moonshot AI provides additional details about the reported activity. Clarification about the training methods, scale of model interactions and contribution of Astra outputs would help establish how the technique influenced the latest models and whether the reported activity breached applicable terms.
More broadly, the AI industry is likely to invest further in protecting proprietary models while refining the rules governing legitimate distillation. As competition intensifies, the ability to learn from existing models will remain technically important, but authorization, transparency and responsible access will increasingly shape how companies use that capability.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
