We introduce a new gradient noise model for stochastic convex optimization and apply it to achieve state-of-the-art rates in both quantum and non-quantum settings.