Since the completion of the EarthScope Transportable Array deployment across the contiguous U.S., there have been various 3-D seismic models with improved image resolution for the crust and upper mantle. However, discrepancies exist between these models due to the differences in both data sets and tomographic methods. It is an essential yet often missing step to evaluate these models and compare their predictability of seismic waveforms, which is helpful for guiding the interpretation and direction of further model refinement. In this work, we systematically evaluate eight 3-D seismic models by measuring the waveform similarity and misfit between observed data and synthetic data computed using these models. An independent validation data set consisting of waveforms from 30 earthquakes recorded by about 3,000 stations across the contiguous U.S., Canada and Mexico is used. The results show that existing 3-D seismic models well capture long-wavelength waveforms while have discrepancy in short-period body and surface waves, suggesting the smaller-scale structures and the radial anisotropy in the crust and upper mantle, especially in the sedimentary basins, are still less constrained. Our results suggest a hybrid model, with the crust constrained by both short-period surface waves (e.g., US.2016) and survey data (CRUST1.0), and with the mantle constrained by unbiased multiple data sets (e.g., S40RTS or S362ANI), can be used as a good initial model for further model improvement based on full waveform inversion.