A model was formulated for estimating dynamic origin-destination (O-D) travel time and flow on a long freeway with a neural Kalman filter originally developed by the authors. The model predicts O-D travel times and flows simultaneously by using traffic detector data such as link traffic volumes, spot speeds, and off-ramp volumes. The model is based on a Kalman filter that consists of two equations: state and measurement. First, the state and measurement equations of the Kalman filter were modified to consider the influence of traffic states for some previous time steps. Then artificial neural network models were integrated with the Kalman filter to enable nonlinear formulations of the state and measurement equations. Finally, a macroscopic traffic flow simulation model was introduced to simulate traffic states on a freeway in advance and predict traffic variables such as O-D travel times, link traffic volumes, spot speeds, and off-ramp volumes. The new model was compared with a regression Kalman filter in which the state and measurement equations are defined by regression models. The numerical analysis indicated that the new model was capable of estimating nonlinearity of dynamic O-D travel time and flow and helped to improve their estimation precision under free-flow traffic states as well as congested flow states.