Abstract The conventional multi‐model ensemble mean usually suppresses the interannual internal variability of El Niño‐Southern Oscillation (ENSO). This study proposed an extended ensemble mean approach for CMIP6 models, which selectively adds or subtracts ensemble members based on their historical correlation with observations. This method significantly improves ENSO variability representation, with correlations of 0.742 and 0.571 against observations at confidence levels of 99% in the training period (1871–2010) and the validation period (2011–2020), respectively. Its projections (2021–2098) reveal a compressed frequency band with two peaks at roughly 4.6 and 3.5 years, and a seasonality of stronger winter variability. La Niña events are projected to be more frequent and persistent than El Niño events. Furthermore, the new ensemble mean projects no strong ENSO events in the next 5 years (2026–2030) and a cluster of two strong El Niño events and two strong La Niña events in the 2040s.