In this note we show the optimal gradient estimate for heat kernels of stable-like operators by providing a counterexample. then there is a smooth fundamental solution p κ (t, x) to the operator L κ satisfying (see [3,4,5])Moreover, p κ (t, x) enjoys the following two-sided estimates: for some K 1 = K 1 (α, d, K 0 ) 1, K −1 1 t(t 1/α + |x|) −d−α p κ (t, x) K 1 t(t 1/α + |x|) −d−α , and gradient estimate: for some K 2 = K 2 (α, d, K 0 ) 1,(1.2)The above estimates can be found in [4,5]. Notice that for λ > 0, if we let κ λ (y) := κ(λ 1/α y), then p κ has the following scaling property: