Anthropic Will Show Users When Fable 5 Refuses Requests or Routes Them to a Weaker Model but Keeps the Restrictions in Place

Anthropic will now show users when Claude Fable 5 refuses a request or routes it to a less capable model. The company said it is changing the safeguards on Fable 5 to make the restrictions visible rather than reversing the policy. Anthropic released Fable 5, a model based on its Mythos system, without disclosing that the model would quietly reroute certain requests.

The restriction applied to tasks including training competing large language models, debugging AI code, and optimizing neural architecture. Researchers discovered the behavior after the model either refused requests or returned degraded responses without explanation in the documentation. Anthropic told Wired that it made the wrong tradeoff on the safeguards.

"We made the wrong tradeoff and we apologize for not getting the balance right," the company stated. It added that if it suspects a user is trying to build a highly capable AI, it will now alert the user that the request is being refused or rerouted. Researchers had raised concerns that the undisclosed limits wasted tokens and money on a model that did not perform as expected.

Research fellow Dean W. " Anthropic has positioned itself as a researcher-friendly alternative to other developers. The company said the new visibility measures address the transparency issues while keeping the underlying safeguard policy in place.