We tested theories of eye movement control in reading by looking at parafoveal processing. According to attention-processing theories, attention shifts towards word n+1 only when processing of the fixated word n is finished, so that attended parafoveal processing does not start until the programming of the saccade programming to word n+1 is initiated (Henderson &, Ferreira, 1990; Morrison, 1984), or even later when the processing of word n takes tao long (Henderson & Ferreira, 1990). Parafoveal preview benefit should be constant whatever the foveal processing load (Morrison, 1984), or should decrease when processing word n outlasts an eye movement programming deadline (Henderson & Ferreira, 1990). By manipulating the frequency and length of the foveal word n and the visibility of the parafoveal word n+1, me replicated the finding that the parafoveal preview benefit is smaller with a low-frequency word in foveal vision. Detailed analyses, however, showed that the eye movement programming deadline hypothesis could not account for this finding which was due not to cases where the low-frequency words n had received a long fixation, but to cases of a short fixations less than 240 msec. In addition, there mas a spill-over effect of word n to word n+1, and there was an element of parallel processing of both words. The results are more in line with parallel processing limited by the extent to which the parafoveal word professing on fixation n can be combined with the foveal n ord processing on fixation n+1.