The distribution of Andean freshwater fishes is the result of the interaction of historical and contemporary factors such as basin geomorphology and the physicochemical characteristics of water bodies. Dramatic changes along river networks due to waterfalls or dams generate abrupt changes in longitudinal slopes that function as ecological barriers to dispersal and thus have an effect on the composition and richness of fish assemblages. We expect the amount of variation in beta diversity along the elevation gradient (between 700 and 3,500 m a.s.l.) of the eastern slope of the Cauca River basin to be explained by changes in the aquatic environment and connectivity among sites. We measured connectivity in terms of the distance along the stream channel between sampling sites considering changes in slope. We used a Generalized Dissimilarity Model (GDM) to evaluate the contribution of connectivity and other water mass characteristics (dissolved oxygen, pH, conductivity, temperature, and elevation) in predicting changes in beta diversity. The GDM models explained 33% of the total deviance in species turnover, suggesting that there are additional variables that have not been considered, such as available habitats along the rivers. Elevation was the variable with the largest relative importance in the model and connectivity explained only seven percent of the total variance when all sites were included. However, when only the sampling sites in the headwater streams were included (the most geographically isolated sites), the GDM models explained 51% of the total deviance and the contribution due to connectivity increased. Isolation of stream headwaters, in conjunction with extreme conditions present at high elevations may influence the fish assemblage turnover. Our results provide evidence that elevation has a strong influence on beta diversity of Andean fish assemblages. Species turnover upstream above 1,200 m a.s.l. is strongly influenced by channel connectivity and there are additional environmental variables that need to be included in the models to improve their prediction power.