Recent advances in automatic differentiation have enabled the development of differentiable simulation environments, accelerating progress across different fields such as robotics and computational physics [1]. These simulators allow gradients of quantities of interest to be computed with respect to controllable inputs, even in the presence of complex physical dynamics, making them well suited for gradient-based optimization. Despite these advances, fluid dynamics has seen limited adoption of differentiable simulation frameworks with control, restricting their use in machine learning-based flow control. In this work, we demonstrate the advantages of integrating differentiable fluid dynamics solvers with reinforcement learning for turbulence control. By exploiting differentiability, our approach improves sample efficiency and uncovers control strategies that are difficult to obtain using traditional, model-free reinforcement learning methods. We validate the method on two representative flow configurations: chaotic two-dimensional Kolmogorov flow, where the objective is to suppress extreme energy dissipation events, and three-dimensional turbulent channel flow, aimed at reducing drag. In both cases, effective control policies are learned with minimal training, highlighting the advantage of gradient-enhanced learning for fluid flow control.