Models of the class used in the CMIP experiments give widely divergent projections in regard to the magnitude of globally averaged warming and especially in regard to regional changes. Averages of the outputs of different models consistently give improved performance as compared to that of any single model, suggesting that model errors due to differences in sub-gridscale parameterizations are effectively random. It has been suggested that inter-model connections in run time could lead to even further improvement. In the resulting interactive ensemble or ''supermodel the constituent models assimilate data from one another using connection strengths for each pair of corresponding variable fields that are determined in a training phase through machine learning. As is now well known in nonlinear dynamics the exchange of small amounts of information between chaotic systems can cause them to synchronize in this case resulting in consensus as to the climate projections. The efficacy of the supermodel scheme has previously been demonstrated with toy models. Here we focus on models of a more realistic type. First we examine synchronization using the intermediate complexity model SPEEDO a primitive equation atmospheric model coupled to the CLIO ocean model. Synchronization is shown to be necessary to avoid loss of variability and to maintain balances. A supermodel is constructed using models with different convection parameters as a proxy for different parameterization schemes. Then we consider a supermodel formed from two versions of the ECHAM model with Nordeng and Tiedtke convection schemes respectively coupled to a common ocean. Although the results indicate that there would be further improvement with direct coupling of the atmospheres coupling through the ocean already gives a much better representation of ocean-atmosphere feedbacks (Bjerknes feedback and heat-flux feedback) than provided by either model separately. Further one is naturally led to consider a range of different weightings of the two atmospheres creating an effective ensemble of supermodels that improves estimates of uncertainty in the “consensus”. The supermodel results are compared to the output averages confirming the hypothesis that with a realistic degree of nonlinearity in the models the supermodel can outperform the ex post facto average in regard to salient higher-order properties of the climate attractor.