The study analyzes the variability of surface air temperature (SAT) and sea surface temperature (SST) obtained from a 1000-yr integration of a coupled atmosphere-ocean-land surface model, which consists of general circulation models of the atmosphere and oceans and a heat and water budget model of land surface. It also explores the role of oceans in maintaining the variability of SAT by comparing the long-term integration of the coupled model with those of two simpler models. They are 1 ) a “mixed layer model,” that is, the general circulation model of the atmosphere combined with a simple slab model of the mixed layer ocean, and 2) a “fixed SST model,” that is, the same atmosphere model overlying seasonally varying, prescribed SST. With the exception of the tropical Pacific, both the coupled and mixed layer models are capable of approximately simulating the standard deviations of observed annual and 5-yr-mean anomalies of local SAT. The standard deviation tends to be larger over continents than over oceans, in agreement with the observations. Over most continental regions, the standard deviations of annual, 5-yr- and 25-yr-mean SATs in the fixed SST model are slightly less than but comparable to the corresponding standard deviations in the coupled mode1, suggesting that a major fraction of low-frequency local SAT variability over continents of the coupled model is generated in situ. Over the continents of both the coupled and the mixed layer models, the spectral density of local SAT is nearly independent of frequency. On the other hand, the spectral density of local SAT over most of the oceans of both models increases very gradually with decreasing frequency apparently influenced by the thermal inertia of mixed layer oceans. However, both SST and SAT spectra in the coupled model are substantially different from those in the mixed layer model near the Denmark Strait and in some regions in the circumpolar ocean of the Southern Hemisphere where water mixes very deeply. In these regions, both SST and SAT are much more persistent in the coupled than in the mixed layer models, and their spectral densities are much larger at multidecadal and/or centennial timescales. It appears significant that not only the coupled model but also the mixed layer model without ocean currents can approximately simulate the power spectrum of observed, global mean SAT at decadal to interdecadal timescales. However, neither model generates a sustained, long-term warming trend of significant magnitude such as that observed since the end of the last century.