Anthropic says rule-breaking customers made up 5% to 7% of its 2026 revenue

Anthropic has told investors something most companies would rather not admit. A meaningful slice of its 2026 revenue came from customers who were breaking its rules. The company behind the Claude models estimates that terms-of-service violators produced 5% to 7% of its revenue across the first nine months of 2026. Most of those violators, according to Anthropic, were Chinese AI labs using Claude to train their own competing models. How the distillation problem worked The practice at the center of this is called model distillation. In AI terms, one lab sends huge volumes of prompts to a stronger model. It then uses the responses as training data for its own system. The result is a cheaper imitation built on someone else’s research spending. Anthropic says the activity ran through thousands of fake accounts. Those accounts generated millions of interactions with Claude to harvest training data in violation of its terms. The scale was not subtle. In February 2026, labs including DeepSeek and Moonshot used around 24,000 fake accounts to carry out more than 16 million interactions with Claude, according to the research findings. Here was the uncomfortable part. Those fake accounts were ...