Description: Poisson regression models are commonly used in actuarial science to model count data e.g., the number of claims and covariates. The standard parametric formulation assumes that the response to covariates is constant across the population under study. Additionally, the shape of the posterior distribution is rather inflexible. In practice, however, the response to covariates may depend on a latent class and the posterior distribution may be highly nonparametric. This paper proposes a Bayesian nonparametric model in which a Dirichlet process prior on the regression coefficients leads to a clustering of observations into groups having the same response to covariates. This paper illustrates the utility of this approach with an example dataset of diabetes deaths.