Graph anomaly detection is a crucial task in graph data analysis, with significant applications in domains such as financial fraud detection and social network analysis. Owing to the prohibitive expense of obtaining labeled anomaly labels, contrastive learning-based anomaly detection methods have been widely adopted. Despite some progress, these methods often struggle to simultaneously capture both local and global features of nodes when constructing sample pairs and lack more effective contrastive mechanisms, which limits their detection performance. Therefore, we propose a framework based on feature fusion and multi-scale contrastive learning (FMC-GAD). FMC-GAD integrates local message passing of Graph Neural Networks and the global attention mechanism of Graph Transformers to achieve feature fusion, effectively capturing both local and global features of nodes. Furthermore, we propose a multi-scale contrastive learning strategy that incorporates local, global, and cross-scale learning, significantly improving anomaly detection capability. Extensive experiments on multiple benchmark datasets demonstrate that FMC-GAD outperforms state-of-the-art methods, validating its effectiveness.