A new approach for the classification of solvents is proposed, based solely on the solvent molecular structure and relying on the “COnductor-like Screening MOdel for Real Solvents” (COSMO-RS), in which solvents are considered in their liquid state. This approach provides an a priori classification without requiring the knowledge of any experimental data. The theoretical descriptors of the proposed classification are generated through the analysis of the COSMO-RS σ-potential profiles. By applying a two-step statistical procedure to the descriptors generated for a set of 153 representative solvents, a clustering into 10 classes leads to an optimal classification that is compared to the classical Chastrette’s classification. The investigation of nitrocellulose solubility allows a validation of the proposed classification, which can be extended also to solvent mixtures. This new approach for solvent classification, thus, appears as a consistent tool for the design of new solvents and their formulation.