We introduce LogitGap, a training-free and post-hoc OOD detector that achieves state-of-the-art performance for both vision-language and vision-only models.