Anthropic Apologizes for Hidden Model Downgrades in Claude Fable 5

After a day of intense community backlash, Anthropic’s newly released model Claude Fable 5 was accused of silently “downgrading” its intelligence whenever the system detected that a user was conducting AI research, effectively making the model less capable without the user’s knowledge.

Anthropic justified the hidden restriction as a measure to prevent foreign adversaries from leveraging the model to accelerate AI development and to protect its competitive edge, but the undisclosed nature of the downgrade sparked widespread criticism.

Under pressure, journalist Max Zeff reported that Anthropic is withdrawing the policy. The company released a statement saying it will make the safety limits for Fable 5 visible to users.

We are rolling out changes to make the safety limits for Fable 5 on frontier LLM development visible.

Starting this week, flagged requests will be clearly rolled back to Opus 4.8, the same safety limits we apply to the network and biology domains. On the API, any flagged request will return the reason for rejection (a server‑side fallback mechanism will be launched in the coming days).

We want to deploy Fable 5 quickly and safely. Visible safety limits can be probed, so they must be robust, which takes time. Invisible limits are more precise and have a very low false‑positive rate, which is why we originally chose them, but that was the wrong trade‑off. We apologize for not balancing this correctly.

Making safety limits visible makes them easier to bypass, so to preserve resistance to jailbreak attacks we will inevitably see more false positives while improving our classifiers. We are adjusting our biology and network classifiers to reduce harmless triggers and will strive to keep this period as short as possible.

If you think a request was mis‑flagged, run /feedback in Claude Code, click the thumbs‑down icon on the fallback prompt at http://Claude.ai or Cowork, or fill out the API safety‑limit appeal form. Your reports help us fine‑tune the classifiers.

The apology has not fully restored user trust; many users remain skeptical that Anthropic might still enforce the policy covertly because the detection mechanisms are hard to verify.

At the same time, competitor OpenAI is pursuing a different strategy by dramatically lowering token prices to win customers. Both companies are preparing for IPOs, and high compute costs remain a shared pain point, potentially leading to unexpected benefits for users as the rivalry intensifies.