A new kind of frequency model reduction method is proposed for unstable processes with time delay based on genetic algorithm. By adopting error between actual model and objective model, the model reduction process is transformed into minimal optimization process. To direct at time-delay problem, phase angle condition is introduced to increase degree of approximation between reduction model and actual model. Performance index is optimized with genetic algorithm to enlarge the applicability of traditional algorithm; the problem of model reduction is solved effectively with this method. Simulation results show that the reduced order model not only can approximate the original model but also can give good dynamic and static characteristics.