Anthropic blocks users for anything touching training larger models

Jiahao Shao, a computer science researcher, accuses Anthropic of restricting his account because it suspected him of using Claude to train a competing model. He says the training rollouts actually came from a powerful open-source model, while Claude was used only to submit jobs and analyze results. Despite that distinction, he says work involving reinforcement fine-tuning, supervised fine-tuning, or training larger models now triggers blocks, ends sessions, or downgrades access from Opus 5.5 to Opus 5. His complaint is that ordinary research automation is being treated as a threat, even by a paying customer.

Shao broadens the dispute into a criticism of AI companies’ rules against large-scale learning from their models’ outputs. He argues that Anthropic and other developers built models using material from the open web, yet prevent users from using model outputs to develop competitors, calling this a double standard that protects commercial advantage. He also contrasts Anthropic’s image as an independent alternative to big technology firms with what he sees as rigid corporate controls imposed on researchers. The central issue is a clash between companies’ efforts to prevent model distillation and researchers’ expectations that paid access permits legitimate experimentation; the account presented is Shao’s, and the claims about why restrictions were applied are his own.