The assessing the accuracy of land cover mappings is essential for the scientific, practical and policy uses of maps. In Brazil, the MapBiomas project has been annually mapping the land cover across the entire territory via automatic classification of Landsat images of medium spatial resolution (30 m) since 1985. Each new version of the classification algorithm generates a new collection of maps that are subject to an accuracy assessment at the national level. However, MapBiomas is increasingly used for regional, municipal or local studies for which the assessment of accuracy at the national level is not adequate. Here we evaluate the accuracy and evolution of accuracy of the main categories of MapBiomas coverage for the state of São Paulo (SP), the most urbanized in the country and the objetic of many studies and public policies related to land cover. We analyzed collections 3.1, 4.1, 5.0, 6.0 and 7.0 for the year 2017, the most recent with coincidence of classes in all collections, considering the classes: Forest Formation, Planted Forest, Pasture, Sugar Cane, Urban Infrastructure and River , lake and ocean. The global accuracy (GA) had its lowest value in collection 4.1 (91%) and highest in 7.0 (96%). Producer (PA) and user accuracies (UA) improved from collection 3.0 to 7.0, with the exception of PA for Planted Forest, which remained virtually unchanged, and for Urban Infrastructure, which has been showing a tendency to worsen over the course of the collections, reaching its lowest value at 7.0 (0,87). Even so, the fact that the GA, and particularly the PA and UA of collection 7.0 are above 0.84 indicates that in Sâo Paulo MapBiomas is accurate enough for, for example, many common analyzes at the landscape scale, in ecology (e.g. distribution modeling of species of mammal or bird), agriculture (e.g. harvest estimation), or engineering (e.g. choice of airport site)