Covariances are crucial for propagating information from data-rich to data-sparse regions, but owing to the sparsity of observations, previous analyses of subsurface ocean temperature and salinity have employed fairly crude parameterizations of covariances defined at resolutions much coarser than the required analysis. Here we investigate the use of covariances computed directly from a realistic coupled global climate model, HadCM3. We demonstrate that 50 to 80% of the monthly variance of both temperature and salinity at 300 m for the period 1950 to 2004 could potentially be captured by analyses based on perfect covariances. However, because of errors in the HadCM3 covariances, we estimate that our analysis captures on average 40% and 25% of the variance of temperature and salinity anomalies in the upper 300 m for this period, although we project both these percentages to increase to about 70% for future analyses benefiting from the full ARGO array. Assessed against tide gauge and altimeter data, our analysis is comparable to that of Ishii et al. (2003) (I03). However, the simple average of our analysis and I03 provides a clear improvement over I03. This improvement probably arises partly from additional observations and partly because errors in covariances will be reduced by taking the mean of an ensemble which samples their uncertainty. We recommend sampling all sources of uncertainty in future ensemble analyses, and anticipate that statistics from future more realistic climate models will become increasingly important to such an ensemble.