In biometrics fusion, the match score level fusion has been frequently adopted because it contains the richest information regarding the input pattern. However, in practice, the size of training match scores increases almost exponentially with respect to the number of users. Under this situation, the cost of learning computation and memory usage can be very high. In this paper, we propose an online learning algorithm to resolve the computational problem. While the existing recursive least squares learning approach contains a mismatch between its objective function and the desired classification performance, the proposed online learning directly optimizes the classification performance with respect to fusion classifier design. Since the proposed method includes a weight that varies according to the class type of newly arrived data, an online learning formulation is non-trivial. Our empirical results on several public domain databases show promising potential in terms of verification accuracy and computational efficiency.