AbstractCoupled Model Intercomparison Project Phase 5 (CMIP5) models have limited capacity for depicting the variability of precipitation at interannual and interdecadal timescales. This study analyses the relative magnitude of interannual to multidecadal variability in annual precipitation across the world in the recently released CMIP6 dataset by quantifying the discrepancy between observed and modelled CMIP5 and CMIP6 data, considering both the absolute and relative magnitude of precipitation variability. We found that: (a) similar to CMIP5, CMIP6 models were short of capacity to simulate the non‐homogeneity in the relative magnitude of interdecadal variability, which is linked to prolonged drought and pluvials in terms of the strength value. The relative value of the interdecadal variability ranged from 15% to more than 30% in observation. Compared with observational data, the relative magnitude, having spatial uniformity, mostly ranged from 10 to 20% in CMIP5 and CMIP6. This result suggests that future projections lack a sufficient decadal variability in CMIP6, indicating a limited capacity for the prediction of floods and droughts in regions like central Africa, North America, and Amazonia; (b) Most individual models in CMIP6 had a better performance in terms of the spatial distribution of the interdecadal precipitation as compared to CMIP5. However, the absolute variations of the overall, interannual, and interdecadal precipitation of multi‐model ensemble simulations (MME) in CMIP6 were larger than those in CMIP5; (c) the underestimation of the interdecadal component in different areas was markedly decreased in East Asia, South Asia but increased in North and South Africa in CMIP6, when compared with CMIP5 and (d) For future scenarios, the higher the range of future forcing pathway is, the larger interdecadal and interannual variabilities are in all regions in CMIP6. For future persistent droughts and floods predicting in a specific region, it is necessary to consider the discrepancy between historical simulations and observational data, and select the performance‐plus models.