Modes of low-frequency circulation variability in the Northern Hemisphere extratropics are compared between five reanalyses. Circulation modes are detected by rotated principal component analysis (PCA) of monthly mean 500 hPa geopotential heights between 1957 and 2002, separately for individual seasons. The quantification of differences between reanalyses is based on the percentage of gridpoints (approximately corresponding to the percentage of area) where the spatial representations of a mode (loadings) significantly differ between reanalyses. The differences between surface-input reanalyses (20CRv2c, ERA-20C) and full-input reanalyses (NCEP-1, ERA-40, JRA-55) are larger than differences within the reanalyses groups in all seasons except for autumn. The causes of the differences are of two kinds. First, the differences may be inherent to PCA, namely, the spatial structure of the modes may be sensitive to the number of components rotated. This concerns only a few modes. Second, the differences may reflect real correlation structures in reanalysis data. We demonstrate that the differences concentrate in three or fewer modes in each season. The reanalysis most different from the rest is 20CRv2c, the differences concentrating over the southern half of Asia and in the subtropical belt over the Pacific and adjacent southwestern North America. 20CRv2c disagrees from other reanalyses there predominantly before 1980s, which points to the impact of insufficient amount of assimilated observations. On the contrary, ERA-20C exhibits a higher agreement with full-input reanalyses, which is why we recommend it for studies of atmospheric circulation over the entire 20th century.