Anthropic is adding a “context lock” to Claude’s hidden reasoning. Encrypted reasoning generated by Fable 5.1 through the API now has to be sent back exactly as it was created, along with the system prompt, tools, and message history from the original call. If any of that preceding context is altered, the API either throws an error or silently drops the reasoning block.
That change is designed to stop model distillation. Researchers had found that although users couldn’t read Claude’s encrypted reasoning, they could feed it back to other Anthropic-compatible models for decoding. An attacker could ask Opus to produce high-quality reasoning, then hand the encrypted block to a less guarded model like Haiku and coax it into reciting Opus’s full chain of thought. It’s not just copying the big model’s answers — you’re taking the scratch paper with the whole problem-solving process, too.
Anthropic already tried plugging this hole once when it released Fable 5 in June, binding encrypted reasoning to the model so it couldn’t be handed to smaller models like Haiku for decryption. Fable 5.1 takes the extra step of binding it to the conversation: even if the model stays the same, change the prompt, tools, or chat history and the old reasoning is invalidated.
The lock isn’t fully switched on yet. It’s being enforced starting with API accounts created after August 31 that call Fable 5.1; older accounts won’t feel it for now. Claude.ai, Claude Code, Cowork, and other Claude models aren’t covered either. Anthropic says future models will ship with these rules enabled by default for all users.