This article performs a systematic study to understand the dependence of the splitting of an initially axisymmetric storm on the various components of the Unified Model (UM) and Weather Research and Forecasting model (WRF). The models are adapted to keep their differences at a minimum. Results at km‐scale grid resolution show that the models differ significantly even under a controlled environment with no surface and radiative forcing. The initial storm in UM splits into two within the first hour. WRF also produces two separate updraughts, but it does not split entirely because of a secondary downdraught that falls just ahead of the original updraught. The cold pool from this downdraught converges with the oncoming winds at lower levels to generate a ring of updraughts connecting the two. UM also shows a similar secondary downdraught, but it is relatively weak. Experiments with the successive reduction in complexity of the microphysics scheme show that the models start to differ with the inclusion of rain processes. Sensitivity experiments with the magnitude of turbulent mixing length do not impact this aspect of model behaviour. Resolution sensitivity experiments show that the storm does not split in UM for a horizontal resolution of O(100 m), whereas WRF behaves consistently across all the resolutions. Through further analyses, we argue that the formal accuracy of the model dynamical core has no control in deciding whether the initial storm will undergo a split or not. The life cycle of a thunderstorm is complex, and to correctly capture its evolution is vital for a numerical model. Here we have used two well‐established models (UM and WRF) to show that even in an idealized set‐up the models differ significantly (see image). Results demonstrate that the models start to differ with the inclusion of rain processes, which points to the physics–dynamics coupling as a potential cause for it.