Global gridded crop models underestimate yield losses from climatic extremes

Global gridded crop models (GGCMs) are crucial for projecting the impacts of climate change, yet their performance under climatic extremes remains poorly understood. Using historical subnational yield data, we evaluate 13 GGCMs from the Inter-Sectoral Impact Model Intercomparison Project (ISIMIP) across major crops. Models largely capture the signal of heat and drought extremes but consistently underestimate the magnitude of losses. For extreme wet conditions, they often miss both the signal and magnitude (losses underestimated in 83–95\% of cases, best to worst model). However, underestimation is reduced for compound heat–drought extremes (38–77\%). Notably, the model ensemble median, widely used to reduce uncertainty, performs poorly under extremes, systematically underestimating risks. We identify model characteristics that influence underestimation under extremes, such as the representation of radiation or soil nutrient dynamics. Improving key process representations and input data quality is essential to strengthen climate risk assessments for global agriculture and food security.