We present a method for testing the performance of global ocean carbon cycle models using measurements of atmospheric O2 and CO2 concentration. We combine these measurements to define a tracer, atmospheric potential oxygen (APO ≈ O2 + CO2), which is conservative with respect to terrestrial photosynthesis and respiration. We then compare observations of APO to the simulations of an atmospheric transport model which uses ocean-model air-sea fluxes and fossil fuel combustion estimates as lower boundary conditions. We present observations of the annual-average concentrations of CO2, O2, and APO at 10 stations in a north-south transect. The observations of APO show a significant interhemispheric gradient decreasing towards the north. We use air-sea CO2, O2, and N2 fluxes from the Princeton ocean biogeochemistry model, the Hamburg model of the ocean carbon cycle, and the Lawrence Livermore ocean biogeochemistry model to drive the TM2 atmospheric transport model. The latitudinal variations in annual-average APO predicted by the combined models are distinctly different from the observations. All three models significantly underestimate the interhemispheric difference in APO, suggesting that they underestimate the net southward transport of the sum of O2 and CO2 in the oceans. Uncertainties in the model-observation comparisons include uncertainties associated with the atmospheric measurements, the atmospheric transport model, and the physical and biological components of the ocean models. Potential deficiencies in the physical components of the ocean models, which have previously been suggested as causes for anomalously large heat fluxes out of the Southern Ocean, may contribute to the discrepancies with the APO observations. These deficiencies include the inadequate parameterization of subgrid-scale isopycnal eddy mixing, a lack of subgrid-scale vertical convection, too much Antarctic sea-ice formation, and an overestimation of vertical diffusivities in the main thermocline.