Low-cost gas multi-sensor devices could be efficiently used for densifying the sparse urban pollution monitoring mesh if equipped with a reliable calibration able to counter specificity and stability issues of solid-state sensors they rely on. In this work, we present a neural calibration for the prediction of benzene concentrations using a gas multi-sensor device (solid-state) designed to monitor urban environment pollution. The feasibility of a sensor fusion algorithm as a calibrating tool for the multi-sensor device is discussed. A Conventional air pollution monitoring station is used to provide reference data. Results are assessed by means of prediction error characterization throughout a 13 months long interval and discussed. Relationship between training length and performances are also investigated. A neural calibration obtained using a small number of measurement days revealed to be capable to limit the absolute prediction error for more than 6th month, after which seasonal influences on prediction capabilities at low-concentrations suggested the need for a further calibration.