This paper shows how two publicly available epidemiological modelling packages, namely the Spectrum AIDS Impact Model and the ASSA2003 AIDS and Demographic Model, predict very different impacts from rolling out highly active antiretroviral treatment (HAART) on new HIV infections. Using South Africa as a case study, it shows that the ASSA2003 model predicts a significant drop in new HIV infections as HAART is rolled out, whereas the Spectrum model assumes that HAART does not have a preventative impact (and in fact generates a small increase in new HIV infections). Users will thus draw different conclusions about the public health benefits of HAART depending on which modelling package they use. Despite being presented as a policy-oriented modelling tool capable of exploring 'what if' questions about the impact of different policy choices, the Spectrum model is illequipped to do so with regard to a HAART rollout. Unlike Spectrum, ASSA2003 is more flexible and its assumptions are clear. Better modelling and more information (including about the relationship between HAART and sexual risk behaviour) is required to develop appropriate public-policy modelling for the HAART era.