We model the carbon balance of European croplands between 1901 and 2000 in response to land use and management changes. The process-based ORCHIDEE-STICS model is applied here in a spatially explicit framework. We reconstructed land cover changes, together with an idealized history of agro-technology. These management parameters include the treatment of straw and stubble residues, application of mineral fertilizers, improvement of cultivar species and tillage. The model is integrated for wheat and maize during the period 1901–2000 forced by climate each 1/2-hour, and by atmospheric CO2, land cover change and agro-technology each year. Several tests are performed to identify the most sensitive agro-technological parameters that control the net biome productivity (NBP) in the 1990s, with NBP equaling for croplands the soil C balance. The current NBP is a small sink of 0.16 t C ha−1 yr−1. The value of NBP per unit area reflects past and current management, and to a minor extent the shrinking areas of arable land consecutive to abandonment during the 20th Century. The uncertainty associated with NBP is large, with a 1-sigma error of 0.18 t C ha−1 yr−1 obtained from a qualitative, but comprehensive budget of various error terms. The NBP uncertainty is dominated by unknown historical agro-technology changes (47%) and model structure (27%), with error in climate forcing playing a minor role. A major improvement to the framework would consist in using a larger number of representative crops. The uncertainty of historical land-use change derived from three different reconstructions, has a surprisingly small effect on NBP (0.01 t C ha−1 yr−1) because cropland area remained stable during the past 20 years in all the tested land use forcing datasets. Regional cross-validation of modeled NBP against soil C inventory measurements shows that our results are consistent with observations, within the uncertainties of both inventories and model. Our estimation of cropland NBP is however likely to be biased towards a sink, given that inventory data from different regions consistently indicate a small source whereas we model a small sink.