User Says Anthropic Blocked Training-Related Work
- •Jiahao Shao said Anthropic blocked his account after flagging his work as “training a competitor.”
- •Shao said an open-source model generated rollouts while Claude submitted jobs and ran analysis.
- •He said RFT or SFT work could trigger blocks, downgrade Opus 5.5 to Opus 5, or end sessions.
Jiahao Shao said on X on October 31, 2026, that Anthropic blocked his account and restricted Claude sessions after its systems flagged his work as “training a competitor.” Shao said he had not distilled Claude: he used a strong open-source model to generate rollouts, while Claude helped submit jobs and run analysis. He said work involving RFT, SFT, or training larger models could trigger a block, downgrade Opus 5.5 to Opus 5, or end a session.
Shao characterized his work as research automation by a paying user and criticized Anthropic’s restrictions. In a follow-up post, he accused the company of applying a double standard by training models on material from the open web while restricting large-scale learning from its outputs. Other commenters on the thread also criticized Anthropic’s restrictions on paying researchers; the discussion had 480 points and 262 comments. The account of the restriction and Shao’s claims come from his posts on X; the thread does not provide Anthropic’s response or an independent account of why the restrictions were applied.