Abstract The representation of the stratosphere and stratosphere–troposphere coupling processes is evaluated in the subseasonal Global Ensemble Forecast System, version 12 (GEFSv12), hindcasts. The GEFSv12 hindcasts develop systematic stratospheric biases with increasing lead time, including a too strong boreal wintertime stratospheric polar vortex. In the tropical stratosphere, the GEFSv12 winds and temperatures associated with the quasi-biennial oscillation (QBO) tend to decay with lead time such that they underestimate the observed amplitudes; consistently, the QBO-associated mean meridional circulation is too weak. The hindcasts predict extreme polar vortex events (including sudden stratospheric warmings and vortex intensifications) about 13–14 days in advance, and extreme lower-stratospheric eddy heat flux events about 6–10 days in advance. However, GEFSv12’s ability to predict these events is likely affected by its zonal-mean circulation biases, which increases the rates of false alarms and missed detections. Nevertheless, GEFSv12 shows stratosphere–troposphere coupling relationships that agree well with reanalysis and other subseasonal forecast systems. For instance, GEFSv12 reproduces reanalysis relationships between polar vortex strength and the Northern Annular Mode in the troposphere. It also exhibits enhanced weeks 3–5 prediction skill of the North Atlantic Oscillation index when initialized during strong and weak polar vortex states compared to neutral states. Furthermore, GEFSv12 shows significant differences in Madden–Julian oscillation (MJO) amplitudes and enhanced MJO predictive skill in week 4 during easterly versus westerly QBO phases, though these results are sensitive to the level used to define the QBO. Our results provide a baseline from which future GEFS updates may be measured.