Gradient vs Derivative-Free

Gradient methods need derivatives; derivative-free methods work on black-box objectives.

When would you choose a derivative-free optimizer over gradient descent?