The discovery of a number of gamma-ray bursts (GRBs) with duration exceeding 1000 s has opened the debate on whether these bursts form a new class of sources, the so-called ultra-long GRBs, or if they are rather the tail of the distribution of the standard long GRB duration. Using the long GRB sample detected by Swift, we investigate the statistical properties of long GRBs and compare them with the ultra-long burst properties. We compute the burst duration of long GRBs using the start epoch of the so-called "steep decay" phase detected with Swift/XRT. We discuss also the differences observed in their spectral properties. We find that ultra-long GRBs are statistically different from the standard long GRBs with typical burst duration less than 100-500 s, for which a Wolf-Rayet star progenitor is usually invoked. Together with the presence of a thermal emission component we interpret this result as indication that the usual long GRB progenitor scenario cannot explain the extreme duration of ultra-long GRBs, their energetics, as well as the mass reservoir and size that can feed the central engine for such a long time.