Anthropic accuses Alibaba of operating a massive model distillation campaign using 25,000 fake accounts to extract and replicate Claude model capabilities into Qwen, demonstrating large-scale IP theft targeting US frontier models.

Anthropic has accused Alibaba's Qwen AI lab of orchestrating the largest known model-copying operation against its Claude system. In a letter sent to US Senators Tim Scott and Elizabeth Warren on June 10, 2026, the company alleged that operators connected to Alibaba created roughly 25,000 fraudulent accounts and generated over 28.8 million exchanges with Claude between April 22 and June 5, 2026—a systematic six-week extraction campaign targeting the model's most advanced capabilities.

Model distillation, the technique at the center of the allegations, involves a smaller or cheaper model querying a more powerful one millions of times to study its responses and learn how it "thinks." The student model then replicates those capabilities without requiring the billions of dollars in training costs. In this case, Anthropic says the operation specifically targeted software engineering and agentic reasoning functionalities associated with its Mythos Preview model.

Alibaba has firmly denied any wrongdoing, calling the accusations baseless. This is not the first time Anthropic has raised alarms about Chinese AI labs copying its work. In February 2026, the company disclosed similar distillation campaigns by other Chinese AI labs, reportedly including DeepSeek and Moonshot.

The competitive pressure is already visible in the market. On platforms like OpenRouter, Chinese models have accounted for between 30% and 58% of token usage among US developers in recent periods. By taking its complaints directly to the Senate and framing model distillation as a national security and intellectual property issue, Anthropic is positioning itself to benefit from any regulatory response. Stricter export controls on AI model access are already being discussed in Washington.

The economics underscore the stakes. If Chinese models are capturing 30% to 58% of developer token consumption on major platforms, the revenue implications for US AI companies are substantial. Lower revenue means less capital for research, which means the distillation problem becomes self-reinforcing. Cheaper copies reduce the original's market share, which reduces funding for the next generation of capabilities, which makes the copies even more competitive by comparison.