Traditional recommendation systems usually use the single type of user-item interaction behavior for recommendation. Multi-behavior recommendation explores multiple relationships between user different interaction (e.g., click, favorite) has been proposed to improve recommendation performance. Most existing multi-behavior recommendation methods are usually insufficient to capture heterogeneous information, which often emphasize the differences between multiple behaviors but ignore the common preferences of user-item interactions. In this paper, we propose a novel Graph Contrastive Learning for Multi-Behavior Recommendation (GCLMBR) framework which exploits the differences and dependencies between different user-item interaction behaviors. Firstly, we leverage a graph convolutional network (GCN) to learn the single-behavior representation from multi-behavior data, and integrate the learned single-behavior representation. Secondly, we apply an enhanced GCN with layer attention mechanism to encode global multi-behavior representation from the heterogeneous graph. The graph contrastive learning task is applied to enhance representation capability of the multi-behavior graph and single-behavior subgraphs. Finally, we adopt a multi-task learning strategy to jointly optimize the learning objectives, and predict the user-item interaction behaviors for recommendation. The empirical results on three real-world datasets show that our proposed GCLMBR outperforms various state-of-the-art baselines.