Anthropic Accuses DeepSeek and MiniMax of Stealing Knowledge from Its Models

Anthropic, the creator of the Claude family of models and one of the key players in the race for AI dominance, has made a sharp statement: according to their data, at least three major Chinese AI developers have systematically "extracted" the outputs of their models to then use them for training and enhancing their own products. Among the named companies are DeepSeek and MiniMax, two of the most notable Chinese startups of recent times. The name of the third developer has not yet been disclosed.

This is about a practice known in the industry as distillation. Its essence is simple: instead of spending hundreds of millions of dollars training a model from scratch on massive datasets, you can use an already powerful model as a "teacher." You send it queries, collect the responses, and then train your own model—typically a more compact one—on these "question-answer" pairs. The result is a system that reproduces a significant portion of the original's capabilities at substantially lower cost. Technically, this is not copying neural network weights or hacking servers, but Anthropic qualifies such actions as violations of its service terms and calls them "unlawful extraction."

Understanding this statement requires context from DeepSeek's meteoric rise, which over the past year has transformed from an obscure startup into one of the most discussed players in the global AI industry. DeepSeek's models have demonstrated impressive benchmark results comparable to leading Western systems, while the company positions itself as a developer achieving efficiency with significantly fewer computational resources. MiniMax, in turn, has attracted investor and user attention through multimodal capabilities and an aggressive international market entry strategy. The successes of both companies raised questions among Western competitors: how do they manage to achieve such results under strict export restrictions on advanced chips imposed by the US administration?

Anthropric's statement provides one possible answer—and this answer is extremely uncomfortable for the entire industry. Distillation as a method is nothing new or secret. It is actively used within Western companies themselves: this is exactly how lightweight versions of models are created—those very "mini" and "lite" variants that run on smartphones and low-power servers.

The problem arises when distillation is conducted not by the model's owner, but by a competitor, essentially parasitizing on someone else's investment in research and computation. Anthropic, by various estimates, has spent billions of dollars training its models—and if a significant portion of this knowledge can be "poured" into a competing product through an API for relatively modest money, this calls into question the very economic model of developing frontier AI systems.

For American regulators and politicians, this statement will serve as yet another argument for tightening control. Already, bills are being discussed in Congress that could limit access to advanced AI models via API for users from certain jurisdictions. The administration is actively expanding export controls, extending them not just to chips, but to software and even model weights. Anthropic's accusations add a concrete case to this discussion, one easily transformed into a political narrative about "stealing American technology."

However, the situation is far from straightforward. Chinese companies will surely contest the accusations, and the legal side of the issue remains murky. API terms of service constitute a contract, not a law, and their violation results in account suspension, not criminal prosecution. Moreover, drawing a clear line between "distillation" and "being inspired by results" is technically extremely difficult. If a researcher reads Claude's responses and, based on the understanding gained, improves the architecture of their own model—is that also distillation? The industry will need to develop a consensus on this question, and until it does, such accusations will remain in a gray area.

What is truly important in this story is that it exposes a fundamental contradiction of the era. Companies like Anthropic simultaneously want the widest possible distribution of their models (this generates revenue and strengthens the ecosystem) and control over exactly how the results of these models are used. These two desires are increasingly difficult to reconcile. Every open API is a potential channel for knowledge leakage. And until reliable technical or legal protection mechanisms appear, the race for AI leadership will be accompanied by ever more acute conflicts between those who create models and those who find ways to extract maximum value from them—without asking permission.