The model skill predicting extreme significant wave heights at 50, 5, and 1-year recurrence intervals is evaluated at buoy locations within several regional wave climates along the US East and West Coasts. Two common third-generation spectral wave models are evaluated, a WAVEWATCH III® (WWIII) model with a grid resolution of 4 minutes, and a Simulating WAves Nearshore (SWAN) model, with a resolution of 200-300 m. Both models are used to generate multi-year hindcasts from which sea state statistics used for wave conditions characterization and classification can be derived and compared to those derived from in situ observations made at National Data Buoy Center (NDBC) stations. Our comparisons generally show underbias for extreme wave heights derived from model hindcasts. This underbias is significantly larger for wave heights derived from WWIII hindcasts than those derived from SWAN hindcasts, primarily due to the coarser resolution of the WWIII model. However, simple linear corrections to compensate for this bias can be applied to improve results that are satisfactory for assessing wave energy project risk and wave energy converter (WEC design.