When a stimulus is presented in repetitive fashion cortical neurons exhibit a large variability in spiking activity, that might represent a source of noise that limits the reliability of sensory systems. Recent findings indicate that the stimulus-driven decline of neural variability (NV) is a widespread phenomenon of cortical responses. To investigate whether during online control of hand movements NV allows for identification of different motor states, such as preparation, execution, or change of motor intention, we have studied its evolution in time in premotor, motor and parietal cortex of macaque monkeys performing reaching movements to visual targets or online corrections of hand trajectory, when the target moved in space. The NV has been measured through normalized variance on spike density functions and Fano factor on spike counts. In all the three areas NV decreased beginning about 50 ms after target onset for about 200 ms. By aligning neural activity to hand movement onset, this decrease started 200 ms before the beginning of hand movement, attaining its minimum at about 150 ms after movement onset. The analysis of the relationships between behavioral variables, such as reaction-time (RT), and the variability of neural signals showed that short RT trials had less NV than long RT trials. Overall the results suggested that NV might be a good predictor of motor preparation, also when a sudden change in motor intention is required in changing environments.