Accurate estimates of soil carbon stocks at the field scale are required to run market-based instruments for soil carbon, but the soil measurements required to make these estimates are expensive. Therefore, efficient sample designs are required. We explored the costs associated with estimating the mean soil carbon stocks within a 68-ha field on the old alluvial soils of the Macquarie River in central-west New South Wales (Red Chromosols or Red Luvisols). The sampling required to achieve a particular degree of accuracy depends upon the variability of soil carbon within the field. We conducted a 100-site geostatistical survey to estimate the variogram of soil carbon. We then used this variogram to consider the efficiency with which simple random and stratified sample designs can achieve a standard error <2 t/ha for the mean carbon stock to 30 cm. The stratifications considered were either purely spatial or based upon auxiliary information such as landform or sensor data. The effectiveness of localised clustering or quadrats within designs was also considered. Formulae were devised to determine the costs of implementing the different designs, based upon our experience from conducting the geostatistical survey. Only weak correlations between carbon stocks and the auxiliary information were evident, and hence the stratifications were largely ineffective. Some benefits of using quadrats were evident, since analytical and field survey costs were reduced. However, the cost (AU$2500) required to achieve the target accuracy is still considerable. The sampled field has complex pedology, and we therefore expect that these costs are larger than average. Similar studies are required to calculate sampling requirements in different locations and to determine whether these requirements can be related to factors such as soil type, parent material, or land management history.