Anthropic has walked back a safety policy for Claude Fable 5 after public criticism, acknowledging that it struck the wrong balance when it quietly limited the model’s responses without telling users.
Fable 5, released days earlier, is the public version of Anthropic’s most capable model family, Mythos. At launch, the model automatically routed questions touching on sensitive areas such as cybersecurity, biology and chemistry to the less powerful Claude Opus 4.8, a measure the company says was meant to prevent misuse.
The policy that drew the most criticism
The decision that caused the sharpest reaction was a different one. Anthropic had initially arranged for Fable 5 to quietly reduce its capabilities when a user asked sensitive questions about building AI models or advanced AI research, without informing the user that anything had changed. Many developers and researchers called the practice opaque and misleading.
What changes now
Following the backlash, Anthropic has revised the policy. The company says that when a request is blocked or routed to Opus 4.8, users will now be told clearly that this has happened and why. API customers will also receive the reason for a refusal or a reroute.
In its statement, Anthropic acknowledged that it had made the wrong call and had not found the right balance.
The company maintains that the safeguards were put in place because of national security concerns, with the aim of keeping advanced AI, cyberattack techniques, biological research and other sensitive technology out of the hands of hostile states and dangerous actors. Reporting on the reversal notes that the underlying restrictions on using Fable 5 for frontier AI development remain in place; what has changed is that users are now told when they apply.
Mythos remains restricted
At the center of the debate is Mythos itself, which security organizations and outside experts describe as one of the most powerful AI systems in existence, particularly in advanced reasoning, cybersecurity and scientific research. Anthropic is not releasing Mythos to the general public. Access is limited to governments, security agencies and approved research organizations.
The episode raises a question that extends beyond Anthropic: how far safety limits on AI models should go, and whether users have a right to know when the system in front of them is not operating at its full capability.
