Measuring T2* relaxation during the course of MR-guided radiotherapy can characterize tumour hypoxia, which is associated with treatment resistance. T2* mapping with radial trajectories allows for efficient coverage of k-space but is susceptible to errors arising from gradient delays. We propose a method that jointly estimates gradient delays and T2* using model-based reconstruction. Using the numerical phantom and the in-vivo prostate data we demonstrated that the proposed approach performs better for different noise levels for both fully sampled and undersampled datasets. This will allow better integration of T2* mapping for hypoxia imaging into an MR-linac treatment work flow.