BehaviorGPT is a decoder-only autoregressive architecture designed for smart agent simulation in autonomous driving. It addresses the limitations of existing simulators by eliminating the traditional separation between "history" and "future," treating each time step as the "current" one. This approach simplifies the model architecture, improves data and parameter efficiency, and enables seamless scaling with data and computation. The Next-Patch Prediction Paradigm (NP3) allows models to reason at the patch level of trajectories, capturing long-range spatial-temporal interactions. BehaviorGPT outperforms state-of-the-art models on the Waymo Sim Agents Benchmark, achieving a realism score of 0.741 and improving the minADE metric to 1.540 with a 91.6% reduction in model parameters. The model uses relative spacetime representations and a triple-attention mechanism to model spatial-temporal interactions among agents. It also employs a next-patch prediction head to generate trajectory patches sequentially, enhancing the realism and accuracy of multi-agent simulations. BehaviorGPT demonstrates superior performance in trajectory prediction, realism, and scalability, making it a robust solution for autonomous driving simulations.BehaviorGPT is a decoder-only autoregressive architecture designed for smart agent simulation in autonomous driving. It addresses the limitations of existing simulators by eliminating the traditional separation between "history" and "future," treating each time step as the "current" one. This approach simplifies the model architecture, improves data and parameter efficiency, and enables seamless scaling with data and computation. The Next-Patch Prediction Paradigm (NP3) allows models to reason at the patch level of trajectories, capturing long-range spatial-temporal interactions. BehaviorGPT outperforms state-of-the-art models on the Waymo Sim Agents Benchmark, achieving a realism score of 0.741 and improving the minADE metric to 1.540 with a 91.6% reduction in model parameters. The model uses relative spacetime representations and a triple-attention mechanism to model spatial-temporal interactions among agents. It also employs a next-patch prediction head to generate trajectory patches sequentially, enhancing the realism and accuracy of multi-agent simulations. BehaviorGPT demonstrates superior performance in trajectory prediction, realism, and scalability, making it a robust solution for autonomous driving simulations.