Soils contain the largest actively-cycling terrestrial carbon pool, which is itself composed of chemically heterogeneous and measurable pools that vary in their persistence. Fundamental uncertainties in terrestrial carbon-climate feedbacks still depend on the timing, sign, and magnitude of the response of soil carbon, and its underlying pools, to environmental change. However, model comparisons typically focus on benchmarking only bulk soil carbon stocks and climatological temperature sensitivities. Underlying microbial and mineral-associated pools, and their response to global change, have received increasing attention among empirical studies, yet data limitations still hinder benchmarking of these pools and processes in models at ecosystem- to global-scales. Here we examined the distribution of carbon within particulate and mineral-associated fractions across an ensemble of global soil biogeochemical models, and compared model estimates to a global database of soil fractions. We found that, while bulk soil carbon stocks were seemingly comparable in magnitude and geographic distribution across the models and observations, the spread in underlying pools was much more pronounced. Indeed, the ensemble of models varied nearly 6-fold in the proportion of carbon in mineral-associated fractions, and the majority of models greatly underestimated mineral-associated carbon stocks compared to the observations. Latitudinal differences between the models resulted in divergent pool-specific climatological temperature sensitivities, with implications on projections to global change scenarios. Our study elucidates key structural and theoretical differences between models that drive divergent soil carbon projections, and clearly highlights the need to benchmark underlying carbon pools, in addition to bulk soil carbon stocks.