OBJECTIVES: To evaluate the cost-effectiveness of docetaxel versus the once weekly paclitaxel regimen (Pac-1w) and the nano-particle bound paclitaxel (Nab-Paclitaxel) from the National Health Service perspective (UK). METHODS: AA Markov model was developed to estimate the total treatment related costs and clinical benefits of taxane treatment for MBC patients over lifetime. Patients moved through health states of progression-free survival, progression and death. In absence of head-to-head clinical trials, indirect comparisons were made using Pac-3w as a common comparator in 3 clinical trials (Jones 2005, Seidman 2004, Gradishar 2005). Adverse events, dose reduction, G-CSF use, and post-failure treatment were modelled. Costs of chemotherapy and utility values for health states and adverse events were obtained from published literature. Costs came from 2005 NHS reference costs and British National Formulary 2006. Nab-paclitaxel's per cycle price was assumed to equal docetaxel's. Costs and benefits were discounted at 3.5%. Probabilistic sensitivity analyses were conducted. RESULTS: Compared to Pac-1w, docetaxel yields higher mean health benefits (2.01 vs 1.73 life years (LYs) and 1.18 vs 1.03 quality-adjusted life-years (QALYs)) and costs less than the generic paclitaxel (£13,518 vs £15,008) therefore permits the NHS to save £5400/LY and £9800/QALY. Against Nab-paclitaxel, docetaxel produced higher LYs (2.01 vs 1.62) and QALYs (1.18 vs 0.96) but at a slightly higher cost (£13,518 vs £11,640). Docetaxel had an incremental cost-effectiveness ratio of £4,900/LYG and £8,600/QALY against Nab-paclitaxel. Given the indirect nature of this analysis, ICERs rely on many assumptions and are sensitive to hazard ratios and to the price of treatments. Confidence intervals around the ICERs are consequently high. CONCLUSION: Based on current evidence docetaxel is cost-saving compared to generic pac-1w, which is used but not approved in the UK, and cost effective compared to Abraxane. Data from ongoing head-to-head clinical studies will be necessary to reduce the wide confidence intervals.