The present study developed several machine learning-based cost models to predict an efficient total economic cost per vehicle revenue-mile of urban public bus transport. The models were trained on a built-in dataset from 269 transit agencies providing urban services in the United States from 2015 to 2019. A feature selection strategy was implemented, finding that, for each proposed model, a subset of features determined a large impact on unit cost. These “core” features included commercial speed, average salary expenses per employee, vehicle productivity, and fleet ownership cost per vehicle. Machine learning techniques outperformed the linear regression method in terms of predictive power and robustness (understood as the dispersion of predictive power measures over the training sets). Based on SHAP values, the sensitivity analyses showed that the proposed models could be used to predict the impact of changes in some critical features on corresponding unit costs. The results may be useful for: (i) introducing regulatory constraints to the allocation of national public resources to local public bus transport services, aimed at minimizing the resources needed to provide a given level of service; (ii) defining the maximum economic compensation required by firms involved in competitive tendering for the allotment of service concessions, or firms with monopoly rights (by political choice and/or local public ownership); and (iii) improving service contract management and design by identifying key cost drivers of transit services.