Earth system modeling has fundamentally contributed to our understanding of past, present and future climate. Regional-scale multidecadal to centennial variability has been identified as a model blind spot, as across general circulation model generations they showed much lower levels of temperature variance than reconstructions, and underpredict regional state-dependency. In this talk I will discuss recent work on closing this gap, what this implies for projections of temperature extremes, and how TERRA aims to improve capacities to project global change impacts.