Computational budget is a severe limitation on the automatic calibration of expensive hydrodynamic and water quality models. To tackle this limitation, the present work formulated a metamodeling-based approach for parameter estimation of such models and assessed the computational gains of this approach compared to a benchmark alternative (a derivative-free optimization method). A response surface proxy of the original model was designed to emulate the behavior of the underlying system, employing Latin hypercube sampling as a strategy for the design of computer experiments and kriging as the technique for the analysis of computer experiments. The response surface proxy of the original model was employed in the automatic fine-tuning of model parameters and, finally, the computational gain over the benchmark alternative was estimated. The metamodeling-based approach was tested in the calibration of the hydrodynamic and water quality models of two water reservoirs. The benchmark alternative analysis indicated that the metamodeling-based approach required 20% to 38% less function evaluations to reach a solution with the same quality compared to the benchmark alternative.