One major factor is not the software, but the functional+basis set that you use. GGA-level functionals (e.g. BP86) are often overbinding while for meta-GGA I am not aware of such a tendency, you would have to check out individual benchmarks. Hybrid functionals (e.g. B3LYP) are fitted to describe certain parameters, e.g. ionization potentials or bond lengths in a test set of molecules or solids, as correct as possible. A high-level functional simultaneously requires a suitable basis set, for B3LYP often a set with triple zeta valence is recommended.