The climate predictions provided by the IPCC for the next decades indicate that many regions over the world will be subject to major changes of their climate conditions, with modifications of temperature and precipitation patterns. GCMs run under various scenarios of CO2 emissions provide simulations of these changes, which then can be used to feed hydrological models to provide estimates of changes in flow regimes. However these hydrological models are calibrated under present-day conditions and their ability to well behave under conditions that were never recorded before is difficult to assess. To which extent may this produce uncertainty in hydrological predictions under climate change? Are our present-day models able to extrapolate to these unknown future conditions? This work proposes a diagnosis of the capacity of hydrological models to cope with climate changes. The objective is to evaluate their extrapolation ability using past observations. The assessment framework derives from the differential split sample test proposed by Klemes (1986), in which the hydrological model is tested on calibration and validation periods with contrasted conditions. Here to enhance differences between calibration and validation periods, we chose non-continuous periods constituted by years with extreme (minimum or maximum) characteristics in terms of precipitation (or temperature). In the available record, we selected independent five-year periods (one made of the five wettest years in the record and the other made of the five driest years) on which we applied the split sample test. This test was made on a large set of French catchments, using two hydrological models. Results give some insights on the extrapolation capacity of hydrological models and on the uncertainty that may be generated in climate change applications by the way models parameters are estimated.