One of the clearest manifestations of ongoing global climate change is the dramatic retreat and thinning of the Arctic sea-ice cover(1). While all state-of-the-art climate models consistently reproduce the sign of these changes, they largely disagree on their magnitude(1-4), the reasons for which remain contentious(3,5-7) . As such, consensual methods to reduce uncertainty in projections are lacking(7). Here, using the CMIP5 ensemble, we propose a process-oriented approach to revisit this issue. We show that intermodel differences in sea-ice loss and, more generally, in simulated sea-ice variability, can be traced to differences in the simulation of seasonal growth and melt. The way these processes are simulated is relatively independent of the complexity of the sea-ice model used, but rather a strong function of the background thickness. The larger role played by thermodynamic processes as sea ice thins(8,9) further suggests that the recent(10) and projected(11) reductions in seaice thickness induce a transition of the Arctic towards a state with enhanced volume seasonality but reduced interannual volume variability and persistence, before summer ice-free conditions eventually occur. These results prompt modelling groups to focus their priorities on the reduction of sea-ice thickness biases.