A simple method is proposed to improve the accuracy in prediction of drug solubility using the COSMO-SAC model. In this method, the data of solubility of drug in a pure solvent is used to determine any error in the COSMO-SAC model for the drug-solvent interactions. These values are then used to correct for the solubility in the mixture of solvents. We have examined this method for the solubility of 33 drug compounds in 127 different solvent mixtures (235 systems, 1955 data points) over a temperature range of 0-50 degrees C and a wide solubility range of 10(-1)-10(-6) (mole fraction scale). The error in the predicted solubility reduces significantly from 442% (original COSMO-SAC) to only 88% (COSMO-SAC with corrections), making it a practical approach for drug development. The accuracy of this method found to be comparable to another predictive method, NRTL-SAC.