Monte Carlo detector simulation plays a central role in high-energy physics, providing the foundation for precision measurements and searches for new phenomena. At the same time, it is extremely computationally demanding. For Run 3, the ATLAS experiment implemented a series of targeted optimizations and tunings to reduce CPU usage while preserving physics accuracy. Collectively, these improvements achieved approximately a factor-of-two speed-up compared to the Run 2 baseline. For Run 4, ATLAS is refactoring the simulation code to streamline the interface with Geant4 and make it more maintainable by removing unnecessary layers. In parallel, the collaboration is exploring a transformative direction for Full Simulation by offloading electromagnetic physics to GPUs and moving toward hybrid CPU-GPU execution models. This effort builds on the progress of the AdePT and Celeritas projects, which provide GPU-accelerated EM physics and navigation capabilities. Importantly, ATLAS is already capable of submitting hybrid GPU simulation jobs on the Grid, marking a key milestone toward operational deployment and production-quality workflows. This contribution presents the current status of these developments, including CPU/GPU performance, physics accuracy, and practical deployment, and outlines the short- and long-term enhancements that will guide the evolution of ATLAS Full Simulation toward Run 4 and beyond.