This paper investigates the shared or independent nature of grammatical gender representations in the bilingual mental lexicon and the role word form similarity (as in the case of cognates) plays in these representations. In a translation task from Greek (L1) to German (L2), nouns that had the same gender in both languages were translated faster than nouns with different genders, but only when the L2 target utterance required computation of gender agreement (adjective + noun). This tendency held for both cognates and noncognates. Unlike noncognates, however, gender-incongruent cognates yielded more errors than gender-congruent cognates. These results are interpreted as evidence for a shared L1-L2 gender system with L2 cognates relying more heavily on the L1 gender value than noncognates. [Research for this article was supported by I.K.Y.]