It's not a 100% accurate calculation, as different tasks will benefit more from speed than it will 'latency' (to a point).
I think the calc is (CL / speed) x 2000 = nanoseconds (and lower is 'better').
So (12/2666)x2000 = 9
Vs
(16/3200)x2000 = 10
So the 2666 @ CL12 would be ever so slight 'faster' - but you're not going to notice that in real-world uses, and in some tasks, you might benefit from the higher speed.