We analyze global climate model predictions of soil temperature (from the Coupled Model Intercomparison Project 5 (CMIP5) database) to assess the models’ representation of current-climate soil thermal dynamics, and their predictions of permafrost thaw during the 21st century. We compare the models’ predictions to observations of active layer thickness, air temperature, and soil temperature, and to theoretically-expected relationships between active layer thickness and air temperature annual mean and seasonal cycle amplitude. Models show a wide range of current permafrost areas, active layer statistics (cumulative distributions, correlations with mean annual air temperature and amplitude of seasonal air temperature cycle), and ability to accurately model the coupling between soil and air temperatures at high latitudes. Many of the between-model differences can be traced to differences in the coupling between either near-surface air and shallow soil temperatures, or between shallow and deeper (1m) soil temperatures, which in turn reflect differences in snow physics and soil hydrology. We compare the models to observational datasets to benchmark the permafrost-relevant physics of the models. The models show a wide range of predictions for permafrost loss: 2-66% for RCP2.6, 15-87% for RCP4.5, and 30-99% for RCP8.5. Normalizing the amount of permafrost loss by the amount of high-latitude warming in the RCP4.5 scenario, the models predict an absolute loss of 1.6 ± 0.7 million km2 permafrost °C−1 high-latitude warming, or a fractional loss of 6-29% °C−1.