- Anthropic announced on June 11, 2026 that it is removing the invisible throttling it added to Claude Fable 5.
The startup said the model launched with hidden guardrails that silently rejected certain high‑risk queries. In a statement, Anthropic wrote, “We are reversing course and will be more transparent about when the restrictions kick in, even if that means Fable refuses more queries.” The change means the system will now return an explicit refusal message instead of a silent cutoff.
This matters because developers and rival labs rely on consistent API behavior to train competing systems. Undisclosed throttling skews benchmark results and forces users to guess why a request failed, eroding trust in Anthropic’s platform.
The move mirrors a broader trend of AI firms exposing safety limits after pushback from the research community.