Transport models for infrastructure investment and operations planning make use of generalised trip cost to predict travel choice decisions. In cities, the most important factors in the generalised cost is trip duration. When calibrating such models to achieve simulation fidelity, observed data such as the choice of destination and means of travel recorded in travel surveys are used in estimating model parameters. Ideally, observed travel durations should also be used in the model estimation. However, in the past it was infeasible to record the actual trip durations to any degree of accuracy in travel surveys. Trip durations derived from a transport network model were commonly assumed to be sufficiently representative. Increasing availability of better recorded trip durations from travel surveys and better modelled trip durations from online mapping present the promise of significant improvements in the fidelity of transport models. As a preamble to adopting such data, we investigate how the best developed recording of actual trip durations from the London Travel Demand Survey compares with the most advanced trip duration modelling from Google Map travel directions API. We find clear discrepancies between the two, with the discrepancies varying systematically for different means and purposes of travel. The magnitude of the discrepancies is greater than can be attributed to randomness or noise. The systematic nature of the discrepancies suggests that transport network modelling even in its advanced form still has a long way to go to represent the observed patterns of behaviour, particularly for non-commuting journeys which account for about 80% of all trips made in cities. Since the discrepancies may create a systematic bias in the model parameters, it is of critical importance to understand them better in future analysis.