What exactly does focal loss down-weight?

For a correctly classified example with target-class probability $p_t$ close to one, what is the primary effect of the focal-loss factor $(1-p_t)^\gamma$ with $\gamma > 0$ ?

It assigns the example a larger class-balancing coefficient regardless of $\alpha$

It reduces that easy example's contribution relative to difficult examples

It increases its gradient so that confidence approaches one more quickly

It converts cross-entropy into a margin-ranking objective