This paper studies the regret arising from learning the price sensitivity parameter κ of liquidity takers in an ergodic version of the Avellaneda–Stoikov market making model. The market maker aims to maximize long-run average reward while managing inventory risk, and must learn κ online. The authors propose a learning algorithm based on a regularized maximum-likelihood estimator (MLE) that updates the estimate of κ each time a market order arrives. They prove that the expected regret, defined as the difference between the optimal reward and the reward obtained by the learning algorithm, is bounded by O(ln²T). This is achieved through two key ingredients: (1) establishing that the ergodic constant (optimal average reward) is twice differentiable with respect to κ, leading to a second-order performance gap when using a misspecified parameter; (2) deriving a high-probability bound of order N^{-1/2} for the MLE error using concentration inequalities for Bernoulli signals. The paper also provides an explicit solution to the ergodic HJB equation and proves existence of an invariant measure for the controlled inventory process. Numerical experiments validate the theoretical regret bound and demonstrate the algorithm's robustness. This work is the first to analyze regret in ergodic control of jump diffusions, contributing to the literature on online learning in continuous-time stochastic control.