Metastatic prostate cancer remains a significant health concern, necessitating the use of accurate prognostic tools to predict patient outcomes. This study aims to develop a novel survival prediction model for patients with metastatic prostate cancer. Data for patients diagnosed with metastatic prostate cancer between 2010 and 2020 were sourced from the Surveillance, Epidemiology, and End Results database. Independent prognostic features were identified using the least absolute shrinkage and selection operator (LASSO)-Cox regression. Survival models, including XGBoost, RandomForestSRC, CoxBoost, SuperPC, and the DeepSurv model, were constructed, followed by an assessment of model predictive accuracy for 1-, 3-, and 5-year overall survival (OS) using the area under the receiver operating characteristic curve (AUC). The optimal model was selected based on the highest AUC, followed by feature importance evaluation and Kaplan–Meier survival analysis. A total of 10,408 eligible patients were enrolled, with 15 prognostic features identified by LASSO-Cox regression. The DeepSurv model showed the best overall performance, with a concordance index of 0.6488 and AUC values of 0.696, 0.685, and 0.720 for 1-, 3-, and 5-year OS, respectively. Gleason score and age were confirmed as core prognostic features across multiple algorithms. Survival analysis showed high-risk patients (risk score > median) had significantly shorter OS times than low-risk patients (risk score ≤ median; p < 0.001). The DeepSurv model achieves high OS prediction accuracy for metastatic prostate cancer, and Gleason score and age are key core prognostic factors for this population. This model enables reliable individualized risk stratification, which can enhance clinical decision-making and improve patient outcomes.