Single-wavelength anomalous diffraction (SAD) phasing is increasingly important in solving de novo protein structures. Direct methods have been proved very efficient in SAD phasing. This paper aims at probing the low-resolution limit of direct-method SAD phasing. Two known proteins TT0570 and Tom70p were used as test samples. Sulfur-SAD data of the protein TT0570 were collected with conventional Cu-Kα source at 0.18 nm resolution. Its truncated subsets respectively at 0.21, 0.30, 0.35 and 0.40 nm resolutions were used in the test. TT0570 Cu-Kα sulfur-SAD data have an expected Bijvoet ratio <|ΔF|>/ ∼ 0.55%. In the 0.21 nm case, a single run of OASIS-DM-ARP/wARP led automatically to a model containing 1178 of the total 1206 residues all docked into the sequence. In 0.30 and 0.35 nm cases, SAD phasing by OASIS-DM led to traceable electron density maps. In the 0.40 nm case, SAD phasing by OASIS-DM resulted in a degraded electron density map, which may be difficult to trace but still contains useful secondary-structure information. Test on real 0.33 nm selenium-SAD data of the protein Tom70p showed that even automatic model building was not successful, the combination of manual tracing and direct-method fragment extension was capable of significantly improving the electron-density map. This provides the possibility of effectively improving the manually built model before structure refinement is performed. © 2008 Chin. Phys. Soc. and IOP Publishing Ltd.