Research in second language (L2) learning often considers one modality only during task completion. It is unclear if L2 performance is as accurate whatever the modality. L2 learning at school is characterized by a predominance of written materials. One might expect written L2 word recognition to be more accurate than spoken one. This modality effect could also depend on L2 proficiency and the presence of cognate items, closer orthographically than phonologically for most language pairs. Two experiments were conducted with 50 intermediate proficiency French–English bilinguals. Experiment 1 highlighted this modality effect on accuracy and a session effect reflecting a benefit from oral to written modality on latency. In Experiment 2, which included both cognate and non-cognate words, modality effect was even stronger for cognate words and cognate effect depended on modality. In both experiments, these effects depend on L2 proficiency. These findings are discussed according to bilingual word recognition models.