Résumé : The market portfolio efficiency remains controversial. This paper develops a new test of portfolio mean-variance efficiency relying on the realistic assumption that all assets are risky. The test is based on the vertical distance of a portfolio from the efficient frontier. Monte Carlo simulations show that our test outperforms the previous mean-variance efficiency tests for large samples since it produces smaller size distortions for comparable power. Our empirical application to the U.S. equity market highlights that the market portfolio is not mean-variance efficient, and so invalidates the zero-beta CAPM.