Implementation of regular maintenance and rehabilitation to roads often requires establishment of work zones, which usually result in closing of one or more lanes available for traffic. Closing of traffic lanes could engender disturbances in traffic flow that ultimately result in a reduction of road capacity and an increase of traffic delay. In this study, a new analytic Neuro-fuzzy model is proposed for estimation of work zone capacity incorporating the learning algorithms of Neural Networks into a Fuzzy logic model. Three different adaptive Neuro-fuzzy models are proposed that in each model some modifications are made to the former model. In the first model, fifteen input variables are used. However, in the second model three of variables, which had low correlation with work zone capacity, are not considered in modeling process and in the third model, some of datasets are eliminated. In the meantime, the empirical model proposed by Kim is selected for comparison of results of models proposed in this study, after updating its coefficients using the datasets used to develop the proposed models. The root-mean-square deviation (RMSE) of the first, second, third, and Kim model for testing datasets are respectively 27%, 15%, 10%, and 16% of average Work Zone capacity. Comparing these relative error percentages reckons that the third Neuro-Fuzzy model, which is developed by using 12 variables, is the most accurate model. The variables used in this model are: Lane width, Heavy vehicle percent, Work zone length, Speed Limit, Number of closed lanes, Number of open lanes, Work intensity, Night-time operation, Continuous operation, Ramp presence, Work zone grade, and Work zone duration.