This article presents a comparative study about the predictive capability of COSMO-SAC and COSMO-RS(Ol) models with respect to excess enthalpies of various binary systems. Results of the group contribution methods UNIFAC and modified UNIFAC (Do) are given for comparison. COSMO calculations were performed using sigma profiles based oil density functional theory (BP and B3LYP). These were generated by the quantum chemical programs Turbomole and Gaussian 03, with the basis sets used for the calculation being triple-xi valence polarization (TZVP) for Bill and 6-311G(d,p) in the case of B3LYP. All results are compared with experimental data stored in the Dortmund Data Bank for 10 85 1 binary data sets (5368 different binary mixtures) with over 165 000 experimental data points from nearly 2000 references and are analyzed according to the types of components in the mixture. As expected, the intensively trained mod. UNIFAC (Do) model performs best in predicting the excess enthalpy of binary systems in most cases, but also COSMO-RS type models often lead to astonishingly good results considering the very small number of parameters used. Overall, relative deviations for a common data set of 30.7% (mod. UNIFAC (Do)) and between 56.7% and 80.8% for the different COSMO-RS flavors were obtained. Whenever experimental data indicated a significant change of the excess enthalpy with temperature, the different COSMO-RS type models failed to reproduce this behavior.