In this work we assess and compare observed atmospheric circulation patterns from ECMWF ERA-40 reanalysis and those simulated by 14 coupled ocean-atmosphere circulation models (GCMs) from IPCC-AR4/CMIP3. Circulation patterns and their clustering in weather types (WTs) are described using observed and simulated 850hPa winds over tropical South America (10S-40S and 50-80W), a neural clustering (Self-Organizing Maps) and classification algorithm (hierarchic agglomerative classification). In order to identify the models that best reproduce the observed atmospheric circulation, several statistic tests are applied. Systematic biases are noticed in most models. The principal one is associated to inaccurate trajectories of the low-level winds that cross the Andes Cordillera instead of being deflected by this obstacle. More, observed and simulated WTs/rainfall relationships are analysed with a special focus on the La Plata and Amazon Basins. Significant biases in these relationships are identified, especially in the north-western part of the Amazon Basin. Finally, models that reasonably reproduce atmospheric circulation, WTs and WTs/rainfall relationships in tropical South America are the following: INGV ECHAM4, CSIRO MK3.0, CNRM CM3.0 and MIROC 3.2 high resolution. The results of this work facilitate choosing suitable models for rainfall prevision and for determining the hydrological impacts of the Climate Change in the La Plata and Amazon Basins.