We evaluate the skills of ocean–sea ice general circulation models involved in the Ocean Modeling Intercomparison Project in simulating the ocean mixed layer depth and its seasonal cycle in the Arctic region. During summer months, all models consistently underestimate the mixed layer depth compared to observational data from the Monthly Isopycnal Mixed layer Ocean Climatology and the Ice Tethered Profilers. In fall and winter, the models exhibit great variability compared to observational data, and inter-model comparison reveals differences up to several tens of meters. We analyze the origin of the fall and winter model biases in ice-covered regions, where the seasonal cycle of the surface salinity and mixed layer depth is strongly influenced by brine rejection resulting from ocean–sea ice interactions. Focusing first on the central Arctic Ocean, defined here as the region north of 80 N, we show that all models simulate more or less the same vertical sea ice mass balance and thus similar salt fluxes into the ocean during sea ice freezing. Furthermore, the model ensemble features a strong relationship between the stratification profile in September and the mixed layer depth at the end of winter. The models whose stratification compares the best to observational data also display the most realistic values of the mixed layer depth at the end of winter. We argue that the discrepancies between models are therefore not so much linked to the surface salt balance but rather to the accuracy with which those models reproduce the vertical salinity profile. In short, a weakly stratified ocean tends to create a deep mixed layer, while strong stratification leads to a shallow mixed layer. To substantiate this conclusion, we apply a simple conceptual model, which simulates the month-to-month evolution of the mixed layer depth using as input the vertical salinity gradients and the surface salt fluxes from general circulation models. Quite surprisingly, this simplified dynamics captures very well the behavior of the general circulation models, emphasizing the role of the different vertical stratification in the control of the mixed layer depth. Furthermore, this interplay may also significantly account for the large mixed layer biases observed in other ice-covered regions of the pan-Arctic seas, even though sea–ice ocean interaction is not the only driver of mixed layer variability in fall and winter there.