Although the ability to process numerical symbols may be considered a special case of more general linguistic abilities, deficits affecting numbers and words are usually interpreted within entirely independent frameworks. We report a patient presenting typical deep dyslexia, as confirmed in a series of word and non-word reading tasks. Moreover, the main features of his deficit extended to arabic numerals. The patient was equally unable to read aloud non-words and unfamiliar numerals, whereas he performed significantly better with real words and familiar arabic numerals such as famous dates or brands of cars. Additionally, familiar numerals and words yielded qualitatively similar errors, as did unfamiliar numerals and non-words. This contrasting performance with familiar and unfamiliar numerals seems incompatible with any single-route model of number reading. It is rather consistent with the existence of two routes for number reading: a 'surface' route mapping any digit string into a word sequence according to language-specific rules; and a 'deep' semantic route functioning only with familiar items that possess a specific lexical entry. We therefore suggest that number reading is architecturally similar to word reading, although these two processes probably rest on functionally and anatomically distinct pathways.