Abstract: Objectives To address the limited online autonomous task allocation and planning capabilities of unmanned surface vehicle (USV) swarms during long-duration area coverage missions in open environments, this paper proposes a real-time decision-making and cooperative control method for USV swarms enabled by large language models (LLMs). The method aims to enhance full-process mission planning and real-time autonomous decision-making capabilities of USV swarms. Methods A hierarchical decomposition framework for long-horizon tasks is introduced. By embedding large language models, the framework enables the USV swarm to dynamically interpret natural language commands and autonomously generate efficient, conflict-free paths. To manage uncertainties in long-duration missions, a multi-level evaluation and feedback mechanism based on dynamic prompt generation is designed, allowing real-time adjustments to task allocation and path planning in response to unexpected events. The proposed method is validated through simulation experiments in a maritime island-patrol scenario, and its performance is compared across different LLMs. Results The experimental results demonstrate that the proposed method achieves an average task success rate exceeding 82% in complex, long-horizon, multi-task scenarios. The success rate for real-time response and handling of abnormal situations exceeds 70%. Conclusions The findings confirm the effectiveness of large language models in enhancing the planning capabilities of USV swarms, providing a safe and reliable approach for accomplishing long-duration missions in dynamic environments.