Chinese AI firms use OpenAI outputs to train rival models

The shift to OpenAI models

Chinese artificial intelligence corporations DeepSeek and Moonshot AI have reportedly targeted OpenAI's GPT-6 Astra to improve their own systems. According to a report by Chosunbiz, these companies used the outputs of Astra to develop their most recent AI releases. This move comes as the industry faces increasing restrictions on how data can be shared or used by competitors.

The report indicates that Moonshot AI utilized GPT-6 Astra during the development of its Kimi K3.1 Pro model, which launched last week. Similarly, DeepSeek is said to have used Astra's outputs while building the V4.1-Flash 1017 model, which was unveiled on the 17th. These actions suggest a strategic pivot toward OpenAI's infrastructure as a primary source for training data.

Previously, these Chinese firms reportedly focused on distilling data from Anthropic's Claude. However, Anthropic recently strengthened its security measures to prevent unauthorized use of its reasoning data. Because OpenAI's models are considered relatively easier to access, they have become the new primary target for these large-scale training efforts.

How distillation works

The process used by these companies is known as distillation. In this technique, a smaller or less complex model is trained using the answers and other outputs generated by a high-performing "teacher" model. By doing this, the smaller model can learn to mimic the reasoning and capabilities of the larger system without needing to be built from the ground up.

Distillation is primarily used as a way to save on the significant time and expense required to develop an in-house AI model. Training a state-of-the-art model from scratch requires massive amounts of computing power and high-quality data. By using a pre-existing model like GPT-6 Astra to provide the answers, a company can shortcut the most difficult parts of the development cycle.

While many companies use distillation internally to refine their own products, it becomes a point of contention when done without permission. The goal is to create a rival product that performs similarly to the original but is cheaper to maintain or more accessible to the end user. This creates a dynamic where the most advanced models effectively act as the blueprints for their competitors.

Security and blocked access

Both OpenAI and Anthropic have publicly stated that Chinese corporations have been distilling their models without authorization. These companies use distillation technology for their own internal purposes but have strict policies against outside firms doing the same. The unauthorized use of their outputs is viewed as a violation of their terms and a risk to their intellectual property.

Anthropic has taken aggressive steps to protect its models from these types of attacks. The company reported that after launching new protection measures earlier this month, the volume of reasoning data pulled from Claude by bad actors dropped by 90%–99%. These defenses make it much harder for outside parties to repeatedly probe the model's internal reasoning processes.

In addition to these technical barriers, Anthropic has identified and permanently blocked tens of thousands of malicious accounts. These accounts were specifically associated with distillation attacks. By cutting off these sources, Anthropic aims to prevent other companies from using its technology to build competing models at a lower cost.

The corporate response

OpenAI has also acknowledged the issue and expressed concern over the unauthorized use of its models. A spokesperson for OpenAI stated, "The company takes this issue very seriously and continues to work toward a resolution." This indicates that the company is looking for ways to prevent its models from being used as training sets for rival products.

The situation highlights a growing arms race between AI developers and the firms trying to replicate their success. As companies like DeepSeek and Moonshot AI release new models, the question of how those models were trained becomes a central point of scrutiny. The industry is currently balancing the need for open access with the need to protect the massive investments made in model development.

For the broader market, this means that the cost of entry for high-end AI is being challenged by clever shortcuts. While the technology allows for faster iteration, it also creates a legal and ethical gray area regarding how much a company can