Hi Cristian, rather than comparing the latency differance of two different FFTs why not compair the latency of just ones output with respect to its own input and then the other. It may give more obvious results. The last time I tested this, the latency from input pre overlap delay compencators to output post delay compencators was exactly 7.5 frames, i.e. an FFT of 1024 would be an overall delay of (1024*7.5) samp.
This may well have change as the newer FFTs have bigger frame sizes and no longer revers the partials every half frame, so it's quite posiable the latency has improved as well.