The strong cooling during the Last Glacial Maximum (LGM, 21 ka BP) provides a rigorous test of climate models' ability to simulate past and future climate changes. We force an atmospheric general circulation model with two recent global LGM sea surface temperature (SST) reconstructions, one suggesting a weak and the other a more pronounced cooling, and compare the simulated land surface temperatures (LSTs) to reconstructed data. Our results do not confirm either SST reconstruction. The cold SST data set leads to good agreement between simulated and observed LSTs at low latitudes, but is systematically too cold at mid-latitudes. The opposite is true for the warm SST data set. Differences between the simulated LSTs are caused by varying land surface albedos, which is lower for the warmer SST reconstruction. The inconsistency between reconstructed and simulated climate points to a potentially significant bias in the proxy reconstructions and/or the climate sensitivity of current climate models.