A novel test for CAPM is presented. In contrast to the traditional models, allowance is made for the possibility that the risk measure, β, to be drawn from two different regimes, e.g. high-risk state and low-risk state. Estimation method is given, empirical results are investigated and specification tests are performed. The hypotheses of two states cannot be rejected. In addition, evidence shows that the data from low-risk state are consistent with CAPM whereas the data from high-risk state violate CAPM.