A supermodel connects different models interactively so that their systematic errors compensate and achieve a model with superior performance. It differs from the standard non-interactive multi-model ensembles (NI), which combines model outputs a-posteriori. We formulate the first supermodel framework for Earth System Models (ESMs) and use data assimilation to synchronise models. The ocean of three ESMs is synchronised every month by assimilating pseudo sea surface temperature (SST) observations generated from them. Discrepancies in grid and resolution are handled by constructing the synthetic pseudo-observations on a common grid. We compare the performance of two supermodel approaches to that of the NI for 1980—2006. In the first (EW), the models are connected to the equal-weight multi-model mean, while in the second (SINGLE), they are connected to a single model. Both versions achieve synchronisation in locations where the ocean drives the climate variability. The time variability of the supermodel multi-model mean SST is reduced compared to the observed variability; most where synchronisation is not achieved and is bounded by NI. The damping is larger in EW than in SINGLE because EW yields additional damping of the variability in the individual models. Hence, under partial synchronisation, the part of variability that is not synchronised gets damped in the multi-model average pseudo-observations, causing a deflation during the assimilation. The SST bias in individual models of EW is reduced compared to that of NI, and so is its multi-model mean in the synchronised regions. The performance of a trained supermodel remains to be tested.