Anthropic announced Claude Fable 5 and Claude Mythos 5 on 9 June. Fable 5 was the broadly available version with stronger safeguards around cyber and biological work. Mythos 5 was reserved for trusted programmes where approved specialists needed fewer restrictions for defensive research.
One model family, two access paths
The split addressed a difficult fact: the same ability that finds and repairs a vulnerability can also help exploit it. Anthropic said its safeguards were deliberately cautious and could sometimes block legitimate requests while the company worked to reduce false positives.
Broad use with stricter high-risk controls.
Vetted organisations and defined defensive purposes.
A legitimate request blocked by a cautious safeguard.
What developers needed to understand
Developers should treat safeguard behaviour as part of model selection. Test ordinary allowed work, document false refusals, provide a safe escalation path, and never design a critical workflow that depends on a public model answering every dual-use request.
Key takeaways
- Users know which model and access tier they have.
- Blocked work has a safe alternative process.
- Sensitive cyber activity is logged.
- Model access is removed when the purpose ends.
The release made access policy visible as a product feature. Capability, price, speed, and safeguards had to be evaluated together because they shaped what the model could reliably do in practice.
