Skilled readers have developed a certain amount of tolerance to variations in the visual form of words (e.g., CAPTCHAs, handwritten text, etc.). To examine how visual distortion affects the mapping from the visual input onto abstract word representations during normal reading, we focused on a single type of distortion: letter rotation. Importantly, two leading neurally inspired models of word recognition (SERIOL model, Whitney, 2001; LCD model, Dehaene et al., 2005) make distinct predictions: Whereas the SERIOL model postulates that the cost of letter rotation increases gradually, the LCD model assumes an all-or-none boundary at around 40° to 45° rotation angle. To examine these predictions in a normal reading scenario, we conducted a parafoveal preview experiment using the gaze-contingent boundary change paradigm. The parafoveal previews were identical or unrelated to the target word. Critically, each preview's individual letters were rotated 15°, 30°, 45°, or 60°. Apart from the parafoveal previews, all text, including target words, was presented with letters in their canonical upright orientation. Results showed that the advantage of the identity preview condition in eye fixation times on the target word decreased progressively by rotation angle. This pattern of results favors the view that the cost of letter rotation during normal reading increases gradually as a function of the angle of rotation.