Background: A trade-off between validity and feasibility is inevitable in decision analytic modelling. In HIV infection, progression rates depend, inter alia, on individual case histories. Variability between individuals can conveniently be modelled using first-order Monte Carlo microsimulation. But this makes concurrent modelling of second-order uncertainty problematic, as probabilistic analysis is more readily applied to an aggregate cohort. Objective: The objectives of this study were to compare first-order Monte Carlo simulation to a cohort approach in modelling the cost-effectiveness of antiretroviral therapy (ART) in HIV infection, and to validate the two models against an existing published model (CEPAC). Methods: Two alternative Markov models were developed. Cost-effectiveness analyses (CEA) were performed with each model in which optimised background therapy (OBT) plus enfuvirtide was compared with OBT alone in treatment-experienced patients, from a US payer perspective using 2003 costs, discounting costs and effects at 3% per annum over a lifetime. Disease progression in both models was dependent on CD4 cell count, viral load set point, treatment, response, age, and opportunistic infections. The simulation model (ARAMIS) used tracker variables to accumulate individual case histories. Transition probabilities depended on this prognostic information and on elapsed time. The cohort model (CM) used more disease states than ARAMIS to model history and time dependency partially. Both models were populated with cohort data from the TORO-1 and TORO-2 trials; as distributions in ARAMIS and stratified according to disease states in the CM. Markov cycle lengths in ARAMIS and the CM were 1 and 3 months respectively. Results: Estimated mean total lifetime costs were as follows. For OBT, ARAMIS: $156,700; CM: $148,500 (CEPAC: $151,000), and for OBT + enfuvirtide ARAMIS: $201,600, CM: $180,100 (CEPAC: $205,900). Mean discounted quality adjusted life expectancy (in QALYs) for OBT was estimated as ARAMIS: 3.60; CM 3.64 (CEPAC: 3.78) and for OBT + enfuvirtide as ARAMIS:4.59 and CM: 4.29 (CEPAC: 4.58). Respective figures for undiscounted, unadjusted life expectancy (years) were ARAMIS: 4.90; CM: 4.77 (CEPAC: 4.64) for OBT and ARAMIS: 6.40; CM: 5.76 (CEPAC: 5.57) for OBT + enfuvirtide. Conclusions: For treatment-experienced patients, the new models generated comparable estimates of life expectancy and costs. These results were comparable to the current "gold standard" among models of ART. The cohort and simulation approaches have individual strengths and limitations. The cohort model minimises computing time and simplifies the incorporation of probabilistic analysis of uncertainty, an increasingly required feature. The microsimulation model allows a fuller set of model inputs and avoids over-simplifying assumptions, yet the use of individual case histories allows a more compact set of Markov states. Microsimulation may have greater face validity among clinicians, though technology assessors may see this as less important than probabilistic sensitivity analysis. Our results apply to relatively advanced disease and treatment-experienced patients.