Rapid environmental changes linked to human-induced increases in atmospheric greenhouse gas concentrations have been observed on a global scale over recent decades. Given the relative certainty of continued change across many earth systems, the information output from climate models is an essential resource for adaptation planning. But in the face of many known modeling deficiencies, how confident can we be in model projections of future climate? It stands to reason that a realistic simulation of the present climate is at least a necessary (but likely not sufficient) requirement for a model’s ability to realistically simulate the climate of the future. Here, I present the results of three studies that evaluate the 20th century performance of global climate models from phase 5 of the Coupled Model Intercomparison Project (CMIP5). The first study examines precipitation, geopotential height, and wind fields from 21 CMIP5 models to determine how well the North American monsoon system (NAMS) is simulated. Models that best capture large-scale circulation patterns at low levels usually have realistic representations of the NAMS, but even the best models poorly represent monsoon retreat. Difficulty in reproducing monsoon retreat results from an inaccurate representation of gradients in low-level geopotential height across the larger region, which causes an unrealistic flux of low-level moisture from the tropics into the NAMS region that extends well into the post-monsoon season. The second study examines the presence and severity of spurious Gibbs-type numerical oscillations across the CMIP5 suite of climate models. The oscillations can appear as unrealistic spatial waves near discontinuities or sharp gradients in global model fields (e.g., orography) and have been a known problem for decades. Multiple methods of oscillation reduction exist; consequently, the oscillations are presumed small in modern climate models and hence are rarely addressed in recent literature. Here we quantify the oscillations in 13 variables from 48 global climate models along a Pacific ocean transect near the Andes. Results show that 48% of nonspectral models and 95% of spectral models have at least one variable with oscillation amplitude as large as, or greater than, atmospheric interannual variability. The third study is an in-depth assessment model simulations of 20th century monthly minimum and maximum surface air temperature over eight US regions, using mean state, trend, and variability bias metrics. Transparent model performance information is provided in the form of model rankings for each bias type. A wide range in model skill is at the regional scale, but no strong relationships are seen between any of the three bias types or between 20th century bias and 21st century projected change. Using our model rankings, two smaller ensembles of models with better performance over the southwestern U.S. are selected, but they result in negligible differences from the all-model ensemble in the average 21st century projected temperature change and model spread. In other words, models of varied quality (and complexity) are projecting very similar changes in temperature, implying that the models are simulating warming for different physical reasons. Despite this result, we suggest that models with smaller 20th century biases have a greater likelihood of being more physically realistic and therefore, more confidence can be placed in their 21st century projections as compared to projections from models that have demonstrably poor skill over the observational period. This type of analysis is essential for responsibly informing climate resilience efforts.