Integrated assessment models (IAMs) have become central tools in global assessments of how to achieve the Paris climate target. But how reliable are the insights that can be drawn from IAMs? This thesis identifies and begins to assess three challenges associated with the use of IAM ensembles and individual IAMs to draw insights on climate mitigation. First, it highlights the importance of model independence for the robustness of insights that can be drawn from IAM ensembles. It develops a method that uses model documentation to construct a model family tree and uses this method to identify likely model dependencies between IAMs in the IPCC’s 5th assessment report (AR5). The analysis shows that the 14 most influential IAMs in AR5 form three branches, the largest of which (including MERGE, MESSAGE, MERGE-ETL, REMIND, WITCH, and BET) is responsible for about half of the scenarios in AR5. The model documentation also indicates that an expanding set of policy questions has incentivised a continuous increase in the detail and scope of IAMs over time. The findings give reason to believe that the diversity of model choices and assumptions included in the AR5 IAM ensemble might be limited. Second, it argues, based on a debate on values in science in philosophy, that the exclusively positive estimates of the cost of mitigation in AR5 are problematic because they don’t capture the full range of cost estimates found in the literature and because the uncertainty of the cost of mitigation is important. A review of the literature reveals that general equilibrium models, which are responsible for all the cost estimates in AR5, can (despite claims to the contrary) be modified to generate net negative costs, but that few of the IAMs in AR5 include mechanisms that typically contribute to net negative costs. It is also found that the model intercomparison studies that are responsible for most of the AR5 cost estimates focused only on aspects that increase the cost of mitigation. Overall, this gives reason to believe that the AR5 IAM ensemble might be biased against net negative mitigation costs. Third, it shows that predictions of climate policy impacts based on the Future Technology Transformations (FTT) simulation model are highly sensitive to a scaling parameter whose correct value is deeply uncertain. This result, which is obtained using Monte Carlo analysis and uniform and independent distributions (around ±50% of default values) for investor discount rates, technology build times, technology lifetimes, learning rates, and the scaling factor in a global sensitivity analysis, shows that the use of diffusion theory to derive technology deployment – which is seen by those who designed FTT to present a unique feature of the model – does not in itself ensure reliable predictions. In fact, the result indicates that predictions from both energy system optimisation models, which are more widely used, and FTT depend on similar unknowns related to future rates of technological change. Based on these three challenges, the thesis concludes, a diversity of model choices and assumptions is crucial for ensuring robust insights and for reflecting important uncertainties associated with IAM research.