Optimizing expensive models is a challenging task in the aeronautical design community. Multi-fidelity optimization proposes to use different levels of fidelity to alleviate the overall computational cost of the optimization. How to know which level of fidelity will be useful for the optimization process? In this work, we propose a methodology to decide which levels of fidelity are useful. We construct a Pareto front for the fidelity levels in terms of cost and accuracy, and then select only the Pareto optimal fidelity levels. All the models considered in this work were considered as black-boxes, so a gradient-free optimizer is used to solve the optimization problems. We chose to use a multi-fidelity Bayesian algorithm in order to reduce the number of function evaluations. We showed the advantage of selecting the optimal Pareto fidelity levels in the multi-fidelity process by comparing the cost to achieve convergence using different scenarios for level selection. To this end, we applied our methodology to several analytical test cases and to a simple drone design problem for which we obtained an improvement of 38 percent improvement over the mono-fidelity algorithm and a 24 percent improvement over the multi-fidelity algorithm with random selection.