Several models have been proposed to relax the molecular clock in order to estimate divergence times. However, it is unclear which model has the best fit to real data, and should therefore be used to perform molecular dating. In particular, we do not know whether rate autocorrelation should be considered, or which prior on divergence times should be used. In this work, we propose a general benchmark of alternative relaxed clock models. We have reimplemented most of the already existing models, including the popular log-normal model, as well as various prior choices for divergence times (birth-death, Dirichlet, uniform), in a common Bayesian statistical framework. We also propose a new autocorrelated model, called the CIR process, with well defined stationary properties. We assess the relative fitness of these models and priors, when applied to three different protein data sets from eukaryotes, vertebrates and mammals, by computing Bayes factors using a numerical method called thermodynamic integration. We find that the two autocorrelated models, CIR and log-normal, have a similar fit, and clearly outperform uncorrelated models on all three datasets. In contrast, the optimal choice for the divergence time prior is more dependent on the data investigated. Altogether, our results provide useful guidelines for model choice in the field of molecular dating, while opening the way to more extensive model comparisons.