The Fluxpart partitioning model was employed to assess transpiration, photosynthesis and evaporation, respiration based exclusively through high-frequency eddy covariance (EC) data. The model was implemented across three sites with two vegetation cover conditions (dense and sparse) during the drought episode between 2012 and 2015 in the Caatinga biome in Northeast Brazil. Fluxpart, an open-source Python software, partitions data by analyzing the flux variance similarity (FVS) relationship and conducting correlation analyses of EC data. The main contribution to the evapotranspiration (ET) process was transpiration (T), representing about 64 and 67% of the amount ET in areas with dense and sparse vegetation cover, respectively. When the Caatinga biome is in a good state of conservation, it behaves as a carbon sink; the net ecosystem exchange average was −6 g C m⁻² day⁻¹ (gram of carbon per square meter daily) for dense vegetation and 3 g C m⁻² day⁻¹ for sparse vegetation. Through ET, vapor pressure deficit (VPD), and gross primary production (GPP) via Fluxpart, an assessment was made concerning the underlying water use efficiency (uWUE) to understand the nonlinear effects of VPD on the carbon-water coupling at the ecosystem scale. The overvalues of 60 g C hPa^0.5 kg⁻¹ H₂O⁻¹ (grams of carbon times square root of hectopascals per kilogram of water) prevailing in the rainy season and below 20 g C hPa^0.5 kg⁻¹ H₂O⁻¹ were mostly in the dry season. uWUE was an important indicator of the ability of the Caatinga biome to optimize water loss and carbon gain according to the water available in the soil-plant-atmosphere system to withstand water stress, especially at dense vegetation covers.