We compare air-sea CO2 exchange in an ensemble of global ocean hindcast models to a suite of observation-based products for 1990-2018. Individual products agree closely with regional fluxes, but individual models vary widely in their regional estimates. Despite their regional divergence, individual models estimate similar global mean fluxes, indicating that significant regional compensation occurs to balance in the global integral. Models diverge most strongly from the observed mean flux in the northern and southern subtropics, despite a strong agreement in seasonality. In the Southern Ocean, models estimate a wide range of both mean and seasonality. The ensemble of observation-based products can be used to select the models that best represent the regionally-resolved mean state of air-sea CO2 exchange. Three models are within 3 sigma of the observed estimates in all regions. With this selected model ensemble, the global mean flux is slightly reduced and its uncertainty is reduced by 35%. Plain Language Summary The ocean plays a crucial role in mitigating climate change by significantly modulating the growth of atmospheric carbon dioxide. Multiple ocean models that are regularly used to estimate the ocean's carbon uptake, all provide consistent estimates as to the magnitude and timing of the globally integrated ocean carbon uptake. However, these same models indicate a wide range of regional air-sea carbon flux patterns. Taking the observational perspective, it has recently been demonstrated that regional flux patterns can be constrained using existing data. Directly comparing the models to observation-based estimates at the regional scale, we find that only one-third of the models are consistent with the observed regional fluxes. Improvements to ocean models are needed to better represent ocean carbon uptake processes. Key Points Seven observation-based products produce consistent estimates of air-sea carbon flux mean and seasonality, both globally and regionally Only three of nine ocean biogeochemical models agree with observed regional mean fluxes; this metric is applied as a selection criteria From only the selected models, global mean flux and seasonality for 1990-2018 is more tightly constrained