We define a "generative leap" exponent which tightly captures the sample complexity for efficiently learning Gaussian multi-index models.