When exploring the visual environment, one uses saccades to shift gaze and fixation to gather spatially and temporally localized information. We propose that the temporal structure of our environment should constrain the temporal allocation of saccades. Here, we probe the possibility of learning to control saccadic latencies in a choice paradigm. Six participants made saccades within 80-300ms following a target horizontally stepping by 10 deg between two fixed locations. For each participant we constructed two classes of latencies, "short" and "long", using the first and last quartiles of the individual baseline distribution (e.g. [80;152]ms and [185;300]ms respectively). We then concurrently reinforced each class in three blocked conditions across about 60 experimental sessions per participant using different reinforcement probabilities such that the relative ratio of reinforcement rates for "short" versus "long" latencies was either 9/1, 1/9, or 1/1. Latency distributions followed the reinforcement conditions: distributions shifted toward the shorter or longer values or became strongly bimodal. Moreover, the relative rates of "short" over "long" latencies matched the relative rates of reinforcers earned for the corresponding latencies (slope up to 0.95), which reveal the ability to choose when to saccade. Our results reveal that learned contingencies considerably affect the allocation of saccades in time and are in line with recent studies on the temporal adjustment of behavior to dynamic environments. This study provides strong evidence for fine operant control of saccadic latency, supporting the hypothesis of a cost-benefit control of saccade latencies.