We evaluated three global models of the coupled carbon-climate system against atmospheric CO2 concentration measured at a network of stations. These three models, HadCM3LC, IPSL-CM2-C, and IPSL-CM4-LOOP, participated in the C4MIP experiment and in various other simulations of the future climate impacts on the land and ocean carbon cycle. A new set of performance metrics is defined and applied to quantify each model's ability to reproduce the global growth rate, the seasonal cycle, the El Niño–Southern Oscillation (ENSO)–forced interannual variability of atmospheric CO2, and the sensitivity to climatic variations. Knowing that the uncertainty on the amplitude, in 2100, of the climate-carbon feedback is mainly due to the uncertainty of the response of the terrestrial biosphere to the climate change, our new metrics primarily target the evaluation of the land parameterization of the carbon cycle. The modeled fluxes are prescribed to the same global atmospheric transport model LMDZ4, and the simulated concentrations are compared to available observations. We found that the IPSL-CM4-LOOP model is best able to reproduce the phase and amplitude of the atmospheric CO2 seasonal cycle in the Northern Hemisphere, while the other two models generally underestimate the seasonal amplitude. This points to some shortcomings in describing the vegetation phenology and heterotropic respiration response to climate. We also found that IPSL-CM2-C produces a climate-driven abnormal source of CO2 to the atmosphere in response to El Niño anomalies. Here a good model performance rests upon a realistic simulation of ENSO-type climate variability and the subsequent tropical carbon cycle response. The three climate models underestimate the sea surface temperature warm anomaly during an El Niño, but HadCM3LC does best in reproducing the interannual CO2 variability. More efforts are needed to further develop metrics for assessing the sensitivity of the carbon cycle to climate change, and this work should now be extended to assess ocean carbon models against observations.