Ensemble learning is a powerful learning approach that combines multiple classifiers to improve prediction accuracy. An important decision while using an ensemble of classifiers is to decide upon a way of combining the prediction of its base classifiers. In this paper, we introduce a novel grading-based algorithm for model combination, which uses cost-sensitive learning in building a meta-learner. This method distinguishes between the grading error of classifying an incorrect prediction as correct, and the other-way-round, and tries to assign appropriate costs to the two types of error in order to improve performance. We study issues in error-sensitive grading, and then with extensive experiments show the empirically effectiveness of this new method in comparison with representative meta-classification techniques.