We prove that for every concept class, the optimal query complexity of agnostic active learning is strictly smaller than the sample complexity of agnostic passive learning.