The Arctic winter-time atmospheric boundary layer often features strong and persistent low-level stability which arises from longwave radiative cooling of the surface during the polar night. This stable stratification results in a positive lapse rate feedback, which is a major contributor to Arctic amplification. A second state, associated with cloudy conditions, with weaker stability and near-zero net surface longwave flux is also observed. Previous work has shown that many CMIP5 climate models fail to realistically represent the cloudy state. In this study, we assess the representation of the Arctic atmospheric boundary layer over sea ice during the winter months in global climate models contributing to the latest phase of the Coupled Model Intercomparison Project (CMIP6). We compare boundary layer process relationships seen in these models to those in surface-based and radiosonde observations collected during the recent MOSAiC (2019-2020) field campaign, alongside the earlier SHEBA (1997-1998) expedition, and from North Pole drifting stations (1955-1991). Here, we show that a majority of CMIP6 models fail to realistically represent the cloudy state over winter Arctic sea ice. Despite this, CMIP6 models have a multi-model mean low-level stability which falls within the range recorded by observational campaigns, and are mostly able to capture the observed dependence of low-level stability on near-surface air temperature and wind speed. As the Arctic warms, CMIP6 models predict a decline of winter low-level stability, with the Central Arctic’s mean stability falling below zero in the multi-model mean state by the end of the century under the SSP2-4.5 emissions scenario.