Key Highlights:
- As reported, DeepSeek and Moonshot AI used outputs from OpenAI’s GPT-6 extra to train their newest AI models.
- OpenAI disclosed a coordinated campaign, involving attempts to redact protected reasoning from its models, with clusters linked to individuals associated with Moonshot AI.
- Anthropic said its guardrails reduced data harvesting and distillation attacks by around 92-99%, highlighting the growing industry response to the practice.
Chinese AI developers DeepSeek and Moonshot are reportedly depended heavily on outputs from OpenAI’s GPT-6 Astra model to train their new artificial intelligence systems. Cited by Reuters on October 28, the report comes as competition between U.S. and Chinese technology companies evolves, with AI developers increasingly focused on how advanced AI models are trained and how their capabilities can be regenerated.
The practice, known as distillation, allows developers to use outputs from a current AI model to help train another system. While the approach can make model development faster and less expensive, leading AI companies have raised concerns about competitors using the models to replicate advanced capabilities. OpenAI and Anthropic have described measures designed to restrict large-scale extraction, particularly when developers attempt to copy capabilities without authorization. These developments have placed model distillation at the center of discussions about intellectual property, AI security, and competition between leading technology companies.
How Are DeepSeek and Moonshot AI Models Linked to Distillation ?
Moonshot AI’s Kimi K3.1 pro model, released the week before the report, and DeepSeek’s upgraded V4.1-Flash 1017 model launched on October 17, benefitted from distilling outputs from OpenAI’s GPT-6 Astra. Distillation has become a point of contention among leading AI developers because it can help companies reproduce capabilities developed by others without bearing the same level of training costs. OpenAI, Anthropic and Google have argued that the practice can allow competitors to create advanced systems swiftly.
Distillation is also a technique used legitimately within AI development. Developers can train a smaller model using responses from a more capable system to create a model that is cost-friendly. OpenAI and Anthropic have both described using distillation in their own model development processes. The controversy’s focal point is the unlawful attempt to retract capabilities from another firm’s proprietary systems rather than the technique itself.
People familiar with the reported efforts told the publication that distillers may use multiple accounts to collect enough outputs while avoiding detection. This creates an enforcement drawback for AI developers, which must distinguish normal customer activity from systematic attempts to reproduce the model’s capabilities. It also noted that Chinese developers have faced resource restrictions, including constraints on access to advanced chips.
DeepSeek, Moonshot used OpenAI’s GPT-6 Astra to train latest models, The Information reports https://t.co/pICutna1Ah
— Reuters (@Reuters) October 29, 2026
Distillation is one approach that may help reduce the cost and time required to develop competitive systems, although it is not the only factor behind the progress of Chinese AI models. Research and engineering improvements in training and inference efficiency also contribute to model development.
OpenAI and Anthropic Step Up Efforts to Restrict Distillation
OpenAI has previously mentioned concerns about coordinated attempts to extract capabilities from its models. The company said it had linked a core part of a distillation campaign that began in July to individuals associated with Moonshot AI.
OpenAI said the activity began on July 1 and includes spikes on July 24 and 25, and it recorded 16,000 requests using a relevant extraction pattern from more than 4,000 users. Its investigation identified relative activity across a bundle of more than 15,000 users, which the company said it had fully disrupted by July 28. OpenAI clarified that the figures describe attempted extractions and do not mean every attempt was successful.
The company said the campaign sought to expose protected reasoning through manipulated model interactions. As per OpenAI, the activity did not involve breaking encryption, compromising a database, or gaining direct access to stored user conversations. Instead, the operators attempted to make information that was normally protected available through model responses. OpenAI said it responded by restricting accounts, strengthening technical controls, and improving monitoring for coordinated activity.
Anthropic’s October 12 revelation explained that its guardrails include systems designed to detect suspicious requests, restrict access to protected reasoning, and make repeated extraction attempts more strenuous. The company also said it had banned tens of thousands of accounts associated with distillation activity. It acknowledged that its measures can occasionally affect legitimate users, illustrating the challenge of tightening security without disrupting normal use.
The reported use of GPT-6 Astra by Chinese AI developers highlights the growing tension between the rapid development of AI systems and efforts by model creators to protect the technology. As companies strengthen guardrails against large-scale extraction of model outputs, distillation is likely to remain a key issue in debates over AI competition training practices and access to advanced capabilities. The extent of that contribution remains unclear, as do the company’s responses to the latest allegations.
The distinction matters because reported use of a model’s output does not by itself establish how much of a rival model’s performance came from distillation.
