Clipping a Gradient by Its Norm

A model with two parameters produces the gradient

$\nabla L = (9,\; 12)$

on this batch. The optimizer uses clipping by norm with threshold $c = 3$: if $\lVert \nabla L \rVert > c$ the gradient is replaced by

$\frac{c}{\lVert \nabla L \rVert}\,\nabla L$

and otherwise it is passed through untouched.

Which gradient does the optimizer actually use for this update?