Grassland models often yield more uncertain outputs than arable crop models due to more complex interactions and the largely undocumented sensitivity of grassland models to environmental factors. The aim of the present study was to assess the impact of single-factor changes in temperature, precipitation, and atmospheric [CO2] on simulated soil water content (SWC), actual evapotranspiration (ET), gross primary production (GPP) and yield biomass, and also to link the sensitivity analysis with experimental results. We employed an unprecedented multi-model framework consisting of seven grassland models at nine sites with different environmental characteristics in Europe and Israel, with two management options at three sites. For warming/cooling and wetting/drying, models showed general consistency in the direction of SWC and ET changes, but less agreement regarding GPP and biomass changes. The simulated responses consistently revealed an overall positive effect of CO2 enrichment on GPP and biomass, while the direction of change differed for SWC and ET. Comparing with single factor experimental manipulations, SWC simulations slightly underestimated the observed effect of warming,while the overall mean model sensitivity for biomass (+7.5%) closely matched the mean response observed with 1–2 ◦C warming (+6.6%). The models exhibited lower sensitivity of SWC to wetting or drying compared to the experiments. The overall mean sensitivity of biomass to drying was -4.3%, contrasting with the mean experimental effect size of -9.6%, which proved to be more realistic than the mean wetting effect (+3.2%, against +38.9% in the field trials). The simulated sensitivity of SWC to CO2 enrichment was markedly underestimated, while the biomass response (+12.0%) closely matched the observations (+17.5%). Although the multi-model averaging did not manifestly improve the realism of the simulations, it ensured a realistic response in the direction of change to varying conditions. The results suggest a paradigm shift in grassland modelling meaning that the usual practice of model optimisation/validation needs to be complemented by a sensitivity analysis following