This paper presents a framework for the determination of an efficient level of simulation model fidelity for flexible manufacturing systems, which will achieve acceptable output accuracy with minimum resources and thereby reduce model building effort and computation time. To this end, we first formally define different levels of model fidelity using building blocks available in object-oriented (O-O) modelling, where an operation at a higher level is either decomposed into more detailed operations or subjected to more constraints at a lower level. In this paper, five models with different fidelities are defined. Then, simulation models that conform to these O-O models are constructed. Using these simulation models, intensive experiments are conducted to examine how the factors that characterize an FMS contribute to the relative errors of outputs from different models. Since no actual systems are considered, the results generated from the most detailed simulation model are used as references. The experimental results are then summarized by regression-based meta-models. In the proposed framework, the most efficient model for a new FMS is identified so that the relative error of a model estimated from the meta-model is closest to the threshold value provided by users. This framework is tested by two sample FMSs, and the initial results look quite promising.