Simulations of Arctic temperatures by 19 general circulation models are examined as part of a diagnostic subproject of the Atmospheric Model Intercomparison Project (AMIP). The forcing of all the models by observed sea surface temperatures and sea ice from a 10-yr period (1979-1988) permits comparative evaluations of the model biases as well as the models' simulations of the interannual variations contained i n the observational data. The models capture the latitudinal and seasonal variability of surface air temperatures in the Arctic, although a cold bias of -3.3°C (std dev = 3.4°C) is apparent over northern Eurasia during spring, espe4cially in the models that do not include vegetative masking of the high-albedo snow. The 19-model mean bias over northern North America is less than 2 degrees C in all seasons. Over the Arctic Ocean, the spring temperatures generally have a warm bias that averages 3.0 (std dev = 2.9°C), although the bias is smaller in the models in which the prescribed albedo of sea ice is highest. For the summer season, correlations between simulated cloudiness and surface air temperatures are negative and statistically significant, but the corresponding correlations for the winter months are small and statistically insignificant. The models without gravity wave drag are generally colder than the other models at the Arctic surface, especially during autumn.