A few years ago, the US National Toxicology Program sponsored an exercise aimed at comparing different prediction approaches for carcinogenicity by challenging them on a common set of chemicals. The exercise was considered to be sufficiently completed when 40 (out of 44) chemicals were actually experimentally tested, and the experimental and estimated carcinogenicity were compared. More recently, the rodent results for the remaining 4 chemicals have been disclosed, making it possible to draw definitive conclusions on the comparative exercise. Having analyzed the first subset of results with multivariate statistical methods, we present here the analysis of the complete set of results. The present analysis also considers aspects (e.g., the complementarity of the different systems in identifying the carcinogens), which had not been investigated previously. The conclusion of this study were: (a) the expansion of the data base from 40 to 44 chemicals did not significantly change the results of the exercise; (b) the structure-activity approaches generated prediction profiles different from those generated by the prediction systems mainly relying on the use of experimental data (in vitro and in vivo); (c) the performance of the predictive systems was generally rather limited; (d) the prediction systems were affected by over sensitivity; they were generally capable of identifying the molecules containing the potentially alerting substructures, but were not so refined as to be able to discriminate between potential and actual carcinogenicity; (e) the combination of the systems into batteries did not permit a significant increase in the performance of the individual methods. The need for, and possible approaches to finely tuning the systems are discussed.