This study assesses the capability of CMIP5 models in representing primary modes of boreal winter extratropical low-frequency variability. Rotated principal component analysis is applied to monthly mean output from historical simulations of 14 plus two variants models and National Centers for Environmental Prediction–National Center for Atmospheric Research reanalyses (NNR) to isolate the leading patterns of variability in 500 hPa height. For each model data set, North Atlantic Oscillation (NAO)-like and Pacific-North American (PNA)-like patterns are identified using pattern correlation analysis (against NNR patterns). The relative pattern correspondence among CMIP5 models and reanalyses is further quantified via cluster analyses of the rotated empirical orthogonal function, NAO-like, and PNA-like patterns, respectively. For both NAO and PNA, 18.8% of the model patterns lie within the same cluster as NNR. Composite structural differences among clusters chiefly consist of (a) spatial displacements of or (b) regional magnitude disparities in the primary anomaly features. While all models replicate the basic aspects of PNA, a small minority of models fails to replicate NAO pattern. Overall, the best performing model is the "GFDL-ESM2G." Interestingly, models having a well-resolved stratosphere generally perform more poorly than those without. Model biases in low-frequency mode structure have important consequences for the representation of associated regional anomalies in surface air temperature and storm track behavior. Those differences among clusters are linked to variations in dynamical structures and their relation to the climatological-mean flow. It is concluded that some state-of-the-art models have important deficiencies in representing low-frequency variability and some of these deficiencies are associated with the failure of models to adequately replicate the observed climatological stationary waves.