Several recent studies indicate that saccades elicited in the absence of a visual target are slower than visually-guided movements of the same size. In addition, we have shown earlier that the slower saccades observed in two different paradigms had more asymmetrical (skewed) velocity profiles. Recently, it has been reported that predictive saccades are also slower. An interesting question, which arises if predictive and visually-guided saccades do have different velocity profiles, is whether the time when the transition occurs can be determined from their dynamic characteristics (peak velocity and skewness) and whether this transition latency can serve as a plausible criterion for distinguishing predictive and visually-guided saccades. To investigate this problem, visually-guided and predictive saccades were elicited by various experimental paradigms in six normal human subjects. Eye movements were measured using the double-magnetic induction method. We found that scatter plots of normalized peak velocity against latency showed an abrupt, small (10-20%) increase at a surprisingly short latency (about 30-70 ms). Furthermore, skewness of the saccadic velocity profile showed a significant drop at comparable latencies. There was a tight correlation between the peak velocity and skewness transition latencies of each subject. Considering the shape of the latency histograms in this and earlier studies, as well as other data, it appears unlikely that these very short transition latencies demarcate the distinction between predictive and fully visually-guided saccades. Instead, we suggest the possibility that the visual stimulus can speed up saccades at an earlier time than it can initiate and guide them. If this is the case, the very short transition latencies (mean: about 50 ms) probably represent the sum of afferent and efferent pure time delays in the system and do not include the time needed for the computation of saccade metrical properties.