Anthropic accuses Alibaba's Qwen of largest-known Claude distillation campaign

Anthropic alleges that operators linked to Alibaba's Qwen lab launched what it calls the largest-yet campaign to clone Claude's software-engineering and agentic-reasoning capabilities. According to the company's June 10 letter to Senators Tim Scott and Elizabeth Warren and White House officials, the activity spanned April 22 to June 5, 2026, used about 25,000 fake accounts, and generated 28.8 million conversations designed to 'distill' frontier behavior — effectively training a cheaper competing model on Claude's outputs.

Distillation here means harvesting a stronger model's responses at scale and using them as synthetic training data for a smaller, cheaper student model. Anthropic argues the scale (tens of millions of targeted queries focused on coding and multi-step reasoning) makes this materially different from casual API misuse, and that it specifically went after the capabilities Anthropic considers its frontier moat in the Mythos line.

The accusation lands squarely in the US-China AI competition. Anthropic frames the episode as evidence both of American frontier labs' lead and of the national-security stakes of protecting model outputs; senators are reportedly drafting measures to sanction Chinese firms that improperly access US AI. Coming the same week the US government gated both Anthropic's Mythos and OpenAI's GPT-5.6 Sol to 'trusted partners,' it reinforces a week-long theme of government-mediated control over who can touch frontier capability.

Skeptics — loudest on Reddit — note the irony of an industry built on scraped web data crying foul over data extraction, and Alibaba has not conceded wrongdoing. Watch for whether the senators' legislation materializes, whether Anthropic pursues legal action, and how Alibaba's Qwen responds to a charge that has already dented its stock.