A report published Wednesday, citing people with direct knowledge of the matter, claims that Chinese AI developers DeepSeek and Moonshot AI relied heavily on OpenAI’s GPT-6 Astra while training their latest models. The models named in the report are Moonshot’s Kimi K3.1 Pro, released last week, and an updated version of DeepSeek’s V4.1-Flash, which launched on 17 October. The report has not been independently verified, and neither company has confirmed the allegation.
At the center of the claims is distillation, a technique that lets AI developers learn from a more capable model without rebuilding every capability from scratch. In a typical distillation pipeline, a developer sends large numbers of prompts to a stronger system, collects the answers, and uses those responses as training data for its own model. Done at scale, the method can help a smaller or lower-cost model reproduce some of an advanced system’s abilities at a fraction of the expense of original training.
The practice is not new, but it has become one of the most sensitive issues in the AI industry. Leading labs in the United States argue that distillation can allow rivals to replicate advanced capabilities without making comparable investments in research, computing power, and safety work. The latest report adds another layer to tensions over how Chinese AI companies are developing models that can compete with systems built by US labs.
Key facts at a glance
- The report names Moonshot’s Kimi K3.1 Pro, released last week, and an updated version of DeepSeek’s V4.1-Flash, launched on 17 October.
- The alleged method is distillation: using a stronger model’s outputs as training material for another system.
- OpenAI previously accused people linked to Moonshot of running a coordinated campaign in July to extract hidden reasoning from its models.
- OpenAI said it detected activity involving more than 15,000 users, shut the campaign down by 28 July, banned the accounts involved, and shared findings with other AI labs.
- Anthropic said new safeguards reduced data collected from its Claude models in distillation attacks by around 90% to 99%.
- Moonshot said Kimi K3.1 Pro outperformed GPT-6 Astra on a coding benchmark; DeepSeek positioned V4.1-Flash as a lower-cost alternative to US models.
- Neither company’s benchmark results have been independently verified.
Why distillation is difficult to police
Distillation occupies an awkward space in AI development. On one hand, it can be a legitimate way to compress a large model into a smaller one, reduce inference costs, or adapt a general system to a narrower task. On the other hand, it can be used to harvest capabilities from a competitor’s model at scale, especially when the target system is accessible through an API or a consumer product.
The challenge for leading AI labs is structural. The same models they offer to users can also serve as training material for competitors looking to catch up. Every API call that produces a useful answer can, in principle, become a data point for another model. That makes abuse detection a cat-and-mouse problem: labs can look for unusual traffic patterns, coordinated account creation, automated prompt generation, or other signs of scraping, but they cannot easily inspect what a customer does with every response.
OpenAI’s earlier accusation against Moonshot illustrates how quickly the issue can escalate. On 30 September, OpenAI said people linked to the Kimi developer had run a coordinated campaign in July to extract the hidden reasoning behind its models’ answers. OpenAI said it detected related activity involving more than 15,000 users, shut down the campaign by 28 July, banned the accounts involved, and shared its findings with other AI labs. The details of that case remain disputed, but the allegation has become a reference point in the broader debate over model access and competition.
The models at the center of the report
Moonshot’s Kimi K3.1 Pro arrived last week with claims of strong performance. According to Moonshot, the model outperformed GPT-6 Astra on a coding benchmark. Benchmark claims are common in the AI industry, but they are also difficult to compare across labs because evaluation setups, prompting methods, and hardware configurations can vary. Independent verification is often limited, especially for newly released models.
DeepSeek’s V4.1-Flash was introduced as a lower-cost alternative to US models. The report refers to an updated version of the model that launched on 17 October. DeepSeek has built a reputation for releasing capable models at aggressive price points, a strategy that has intensified pressure on larger AI providers. If the distillation allegations are accurate, they would suggest that some of that cost advantage may have been achieved by learning from a more advanced system rather than through fully independent training.
Both companies have posted strong results in their own evaluations. Neither company’s results have been independently verified, and the reported training methods have not been confirmed. That leaves the claims in the realm of allegation, even as they feed into a wider policy debate about how AI capabilities spread across borders.
A wider fight over access and competition
OpenAI, Anthropic, and Google have all argued that distillation can allow rivals to replicate advanced capabilities without making comparable investments in training. US agencies have also identified Chinese firms they say targeted American AI models. The concern is not only commercial. Advanced AI systems can have national-security implications, and the ability to copy or approximate their capabilities quickly could change the balance of power among labs and countries.
The labs are trying to make distillation harder. Earlier this month, Anthropic said new safeguards had reduced the amount of data collected from its Claude models in distillation attacks by around 90% to 99%. Such safeguards can include traffic monitoring, rate limits, account verification, and detection of automated prompting patterns. But defenders of open access argue that heavy restrictions can also hurt legitimate research, startups, and developers who rely on APIs to build products.
The policy environment is tightening in other ways as well. Governments are considering export controls, model access rules, and security reviews for advanced AI systems. Companies are being asked to document training data, evaluate model behavior, and monitor for misuse. Distillation sits at the intersection of all these issues because it involves both data provenance and competitive fairness.
What the report means for the industry
For leading AI labs, the reported use of GPT-6 Astra highlights a difficult problem: the same models they offer to users can also serve as training material for competitors looking to catch up. Labs want to protect their investments in research and compute, but they also want to sell access to their models. The more capable a model is, the more valuable it becomes both as a product and as a target for extraction.
For Chinese AI companies, the allegations add to an already complicated operating environment. They face export restrictions on advanced chips, scrutiny from US regulators, and skepticism about benchmark claims. At the same time, they have shown an ability to release competitive models quickly and at lower prices. The report, if confirmed, would raise questions about how much of that progress depends on access to foreign systems rather than independent innovation.
For users and enterprises, the dispute matters because it affects trust, pricing, and long-term reliability. If distillation is widespread, customers may benefit from cheaper models that perform well on common tasks. But they may also face greater uncertainty about model provenance, safety guarantees, and the durability of vendor claims. Companies building on AI APIs increasingly need to ask where a model’s capabilities came from and whether those capabilities were obtained through legitimate means.
The report is also a reminder that the AI industry’s competitive advantage is not only about building the best model. It is about controlling access to it. Leading labs are investing in safeguards, monitoring, and legal terms that restrict how their outputs can be used. Rivals are looking for ways around those restrictions. Regulators are trying to draw lines. The result is a fast-moving contest that is likely to shape the next phase of AI development.
For now, the central claims remain unverified. Moonshot and DeepSeek have not confirmed the reported use of GPT-6 Astra, and the named benchmark results have not been independently checked. But the allegations have already added to the debate over distillation, model access, and the cost of catching up in advanced AI.