In this paper we develop a new post-pruning algorithm. This new pruning algorithm uses two or more post-pruning algorithms to prune a decision tree that has been built on training set by different orders, and the “best” tree is selected based either on separate test set accuracy or cross-validations from trees coming from result of the above step. The algorithm is theoretically based on occam's razor that is a simpler model is chosen if two models have the same performance on the training set. An experiment is implemented on three databases in UCI machine learning repository and the new algorithm is employed to compares with two well-known post-pruning algorithms. The results show that the hybrid pruning algorithm effectively reduces the complexity of decision trees without sacrificing accuracy.