Parametric AFT models require the specification of appropriate distribution for the event time, which is often difficult to identify in real-life studies. Methods that are more robust with respect to event time distribution specification are desirable. We develop a new flexible AFT model that does not need the specification of the parametric family of event time distribution. The baseline hazard function is modelled by regression B-splines and thus allows for the estimation of arbitrary shapes. In comprehensive simulations, we validate the performance of our approach in terms of effect estimates, baseline hazard and survival probabilities, and compare with the results from parametric AFT models and the approach proposed by Komarek. The survival probabilities estimated by parametric AFT models with mis-specified event time distribution deviated from the truth. Both the proposed flexible AFT model and the approach of Komarek provided unbiased effects estimates and unbiased survival curves for a variety of scenarios in which the event time follow different distributions. Moreover, the proposed flexible AFT model always yielded more stable estimates of the hazard function.