Visual word recognition studies commonly measure the orthographic similarity of words using Coltheart's orthographic neighborhood size metric (ON). Although ON reliably predicts behavioral variability in many lexical tasks, its utility is inherently limited by its relatively restrictive definition. In the present article, we introduce a new measure of orthographic similarity generated using a standard computer science metric of string similarity (Levenshtein distance). Unlike ON, the new measure-named orthographic Levenshtein distance 20 (OLD20)-incorporates comparisons between all pairs of words in the lexicon, including words of different lengths. We demonstrate that OLD20 provides significant advantages over ON in predicting both lexical decision and pronunciation performance in three large data sets. Moreover, OLD20 interacts more strongly with word frequency and shows stronger effects of neighborhood frequency than does ON. The discussion section focuses on the implications of these results for models of visual word recognition.