Claude Fable 5 is back, running inside the agentic tools your teams already use. Its safeguards shape how likely the model is to misbehave. They do not decide what an agent can reach once it acts. That gap belongs to you, and it lives at the execution layer, not inside the model.