In view of Type I and Type II censored for exponential life distribution cases, the Mazzuchi-Soyer reliability growth model was extended to cover the life test. The Dirichlet distribution was taken as prior distribution in the model. The historical information and expert information were synthetically used. Combined with the life test data of each development stage, the joint posterior distribution of each stage reliability was presented. Then, the Gibbs sampling algorithm was used to compute the posterior inference. The Bayesian estimators and Bayesian lower bound were gained for each stage reliability. Finally, the example shows that the Bayesian model has apparent advantages.