We investigate why similar mean state changes lead to very different ENSO amplitude changes simulated in two coupled general circulation models (CGCMs): the Meteorological Research Institute (MRI) and Geophysical Fluid Dynamics Laboratory (GFDL) models. The analysis of the control and 4xCO2 simulations of the GFDL model indicates that the tropical Pacific SSTA distribution is Gaussian; however, the MRI model has a skewed distribution with increased probabilities for extreme warm events. The skewed distribution in the MRI model suggests the importance of non-linearities in the ENSO physics, whereas the GFDL model lies in the linear regime. Consistent with these differences in ENSO regime, the GFDL is insensitive to the mean state changes, whereas the MRI model is sensitive to the mean state changes associated with the 4xCO2 scenario. Similarly, the low frequency modulation of ENSO amplitude in the GFDL model is related to atmospheric stochastic forcing, but in the MRI model the amplitude modulation is insensitive to the noise forcing. These results suggest that the understanding of changes in ENSO statistics among various climate change projections is highly dependent on whether the model ENSO is in the linear or non-linear regime.