Predicting Origin-Destination (OD) flow is a critical issue in the construction and management of smart cities, with challenges mainly stemming from the complex spatial-temporal dependencies in urban environments for three reasons: first, a considerable number of city grids are interrelated; second, the diversity of travel modes leads to highly imbalanced feature distributions in crowd flow data; lastly, crowd flow and geographical indicators have potential correlations. Inspired by natural language processing, we apply the concept of Generative Pre-training (GPT) to OD flow prediction. The proposed OD-GPT model frames grid sequence prediction as next token prediction in language models. Extensive testing on large-scale mobile signaling datasets from two real-world cities demonstrates the effectiveness of the model. In comparison to baseline models, it exhibits a notable enhancement in one-step prediction accuracy. Furthermore, we explore the model’s multi-step prediction capability and research its adaptability to unconventional day through fine-tuning. The analysis indicates that our model can extract latent geographical features from OD data and generate reasonable embeddings with precise dynamic spatial-temporal dependencies.