It is of fundamental importance to evaluate the ability of climate models to capture the large-scale atmospheric circulation patterns. In the context of a rapidly changing climate, it is equally crucial to quantify the robustness of the modeled changes in the large-scale atmospheric dynamics. Here we approach this problem from an innovative point of view based on dynamical systems theory. We characterize the atmospheric circulation over the North Atlantic in the CMIP5 historical simulations (1851 to 2000) in terms of two instantaneous metrics: local dimension of the attractor and stability of trajectories. We then use these metrics to compare the models to the 20CR reanalysis over the same historical period. The comparison suggests that: i) all the models capture the mean attractor properties and models with finer grids perform better; ii) extremes of the dynamical systems metrics match the same large-scale patterns in most of the models; iii) changes in the attractor properties observed for the 20CR reanalysis - studied by dividing the 1851-2000 period into 3 sub-periods of 50 years each - are not reproduced by the models; iv) some models present significant changes in the dynamical systems metrics over time but there is no agreement on the direction and on the intensity of the shifts