To explore the impacts of hydrometeor radiative effects over subtropical and tropical Pacific and Atlantic Oceans, we quantify the mean radiation biases in historical climate simulations based on how frozen-hydrometeors radiative properties are calculated in CMIP6 models. CMIP6 models are divided with cloud ice only (NOS), with combined (SON1), and with separate treatments (SON2) of cloud ice and falling ice (snow) radiative properties. Over the deep convective regions, NOS models overestimate outgoing longwave radiation (RLUT) and surface shortwave irradiance (RSDS), while underestimate top-of-atmosphere reflected shortwave radiation (RSUT). SON2 models reduce these biases by 4–14 W m−2. However, this improvement is not seen in SON1 against NOS. Spatially averaged absolute biases in radiative fluxes for SON1 models are larger than those of NOS, suggesting that the SON1 approach of falling ice radiative effects may not produce the expected hydrometeor–radiation interactions. Over the south Pacific trade-wind regions, both SON2 and SON1 show similar improvements in RLUT, RSUT, and RSDS with positive absolute bias differences up to 20 W m−2 against NOS, leading to improvement of CMIP6 over CMIP5 ensembles. The seasonal cycles are consistent with the annual means over these two regions except with larger differences between subsets of models during January–May than during June–December. In general, improvement from CMIP5 to CMIP6 due to more participating SON2 models is limited because of offset by SON1. These results suggest that a separate treatment of frozen-hydrometeor radiative properties may be critical for reducing the spread of CMIP models.