Eighteen years of global climate model ensemble simulations have been performed with observed SST to assess the ability of dynamical models to predict seasonal-interannual climate variations. In addition, test suites have been designed to assess the role in climate predictability of land-surface initial conditions, and systematic errors of precipitation and radiation fluxes at the land surface. Remarkably high skill in the simulation of interannual variations in surface temperature are found over much of the globe and North America. Much of this skill appears to be attributable to the influence of SST anomalies, but comparison with long AMIP and C20C simulations shows that the land surface state is an important contributor to the skill. Climate drift in the coupled land-atmosphere model over multi-year time scales clearly degrades the ability of the model to simulate interannual variability. Precipitation skill is much lower, with evidence that initial conditions and the land surface play a greater role in predictability of the water cycle. Correction of systematic errors, by assimilation of water and energy fluxes from atmosphere to land, manifest themselves through the model fluxes from land to atmosphere, improving the simulation of precipitation and temperature. However, significant errors remain, suggesting that the parameterizations of atmospheric physics (convection, PBL, radiation and clouds) remain the weak link in the simulation of the coupled land-atmosphere climate system. Overall, there is a strong indication that with the proper treatment of the land surface, climate anomalies during the boreal warm season may be as predictable as the winter anomalies associated with El Nino.