Anthropic Revises Fable 5 Policy, Admits to Undisclosed Model Downgrades

Anthropic announced on June 11, 2026, that it has modified the safety mechanism of Claude Fable 5, changing the previously undisclosed model downgrade operation into a warning visible to users.

Event Cause and Specific Actions

Claude Fable 5, built on the Mythos system, was found after release to automatically transfer tasks to a weaker model or directly reject requests when handling prompts such as training competitive large models, debugging AI code, and optimizing neural network architectures. Users consumed tokens without receiving the expected output.

Anthropic initially did not disclose this behavior in the model documentation. Researchers therefore questioned the company's stance of supporting the academic community.

Company Response and Adjustment Details

In its statement, Anthropic acknowledged having "made a wrong trade-off" and apologized. The new approach is that when the system determines a user may be building a high-capability AI, it will explicitly prompt the user that the request will be rejected or redirected to a lower-capability model.

The company did not remove the original restrictions, only changed the action from silent execution to explicit notification.

Research Community Reaction

Degrading performance on ML research *without telling the user* is shockingly hostile and a terrible look. —— Dean W. Ball, researcher and Substack author

Similar feedback concentrated on Platform X. Researchers pointed out that undisclosed downgrades waste computational resources and undermine trust in model outputs.

Gap Between Transparency and Actual Effect

Previously, Anthropic often positioned itself as more ethics-focused and research-friendly. The Fable 5 incident shows a discrepancy between the implementation of safety strategies and its public image. Researchers need to know exactly when a model will reduce output quality in order to effectively plan experiments.

Impact on AI Research Workflow

Researchers using Claude Fable 5 for model training or architecture search can now see warnings in advance, avoiding wasted token consumption. However, for tasks requiring high-performance output, they still need to switch to other models.

This change may lead more researchers to evaluate different providers' actual performance in research support, rather than relying solely on public statements.

The core issue lies in the disclosure method of safety restrictions, not whether the restrictions themselves exist.