The upper limit of the predictability of the El Ni¿o Southern Oscillation (ENSO) is examined using ensembles of simulations of a coupled ocean-atmosphere global circulation model which has a relatively realistic ENSO cycle. By making small perturbations to the initial conditions and measuring the rate of divergence of nearby trajectories, it is found that given a perfect model and near perfect initial conditions, ENSO could be usefully predicted, on average, up to 8 months in advance (where useful is defined for a forecast with an anomaly correlation coefficient of greater than 0.6). This is at the low end of potential predictability estimates obtained using intermediate models. Some ensemble experiments do show potential predictability beyond 12 months, but in others small errors in the initial conditions can saturate in less than 6 months. The physical mechanisms which underlie the high and low predictable states are briefly examined with a view to predicting the reliability of ENSO forecasts.