The aim of this paper was to measure the reproducibility of osseous landmark identification from two recently described three-dimensional (3D) cephalometric analyses: 3D-ACRO and 3D-Swennen analyses. The study population consisted of 13 patients examined with spiral 3D computed tomography (CT). We used a previously validated low-dose CT protocol. For each analysis, 22 cephalometric reference landmarks were identified on 3D CT surface renderings. Forty-four reference landmarks were identified per patient. Two series of identifications were performed by two independent observers. In total, 3432 imaging measurements were completed. The intra-observer reconstructed mean log was 1.210+/-1.042mm for the 3D-ACRO analysis, and 1.311+/-1.042mm for 3D-Swennen analysis (comparison: p=0.17 NS). The inter-observer reconstructed mean log was 1.799+/-1.037mm for the 3D-ACRO analysis, and 2.465+/-1.036mm for 3D-Swennen analysis (comparison: p=0.000000002). The difference between the intra- and inter-observer reconstructed mean logs were 1.486+/-1.057mm for 3D-ACRO and 1.880+/-1.056mm for 3D-Swennen analysis. In conclusions: 3D-ACRO analysis was significantly more reproducible than 3D-Swennen analysis (p=0.0027) due to the use of a majority of highly reproducible cephalometric landmarks. Finally, we propose a classification scheme and exclusion criteria for reference landmarks used in 3D cephalometrics, based on inter-observer reproducibility and anatomical reality.