Global Gridded Crop Models (GGCMs) have been widely used to simulate the impacts of global warming on crop production, but their accuracy in capturing the real-world temperature sensitivity of crop yields remains unclear. Here, we evaluated the performance of eight GGCM emulators (incorporating versus not incorporating cultivar adaptation of crop growing periods at 0.5° × 0.5° resolution) in modelling yield sensitivities to 1 K temperature increase (ST) and optimized their ensembles against statistically-inferred ST for maize, rice, and wheat using a Bayesian Model Averaging approach. Our results suggest that multi-GGCM ensembles assuming a fixed crop growing period (i.e., a gradually temperature-adapted crop cultivar) show higher goodness-of-fit to statistically-inferred ST than those assuming a temperature-sensitive growing period for the crops in major food-producing countries. When setting a temperature-adapted growing period instead of a temperature-sensitive growing period in the GGCM ensembles, the R2 between GGCM-simulated and statistically-inferred ST increased from 0.63 to 0.81 for maize, 0.28 to 0.52 for rice, and 0.40 to 0.85 for wheat, meanwhile the RMSE was reduced for all three crops across their respective top 20 producing countries. The crop models may exaggerate historical responses of crop growing periods to climate warming, resulting in an overestimation of yield ST for maize and an underestimation of yield ST for rice and wheat in major food-producing countries. The study highlights the importance of adopting dynamic phenological parameters in GGCM simulations to reflect crop cycle adaptation under global warming.