A set of 44-year seasonal ensemble coupled model forecasts performed with annually updated greenhouse gas concentrations is compared to a standard seasonal ensemble forecast experiment performed with fixed concentrations. The former shows more realistic temperature variability and better forecast quality. The improvement in model variability is due to a better simulation of climate trends and suggests that realistic initial conditions are not enough to reproduce this long-term variability. The better probabilistic forecast quality is mostly due to the increased ability to reliably discriminate the occurrence of events and non-events. These results are relevant for the improvement of operational seasonal forecasts and provide new evidence of the effects of anthropogenic changes in atmospheric composition.