AMIP has been a useful start toward the adoption of community-wide standards for the documentation, analysis and intercomparison of global atmospheric climate models and for the identification of their systematic errors. A preliminary analysis of the zonally-averaged results of the 30 participating AMIP models shows that: (1) Most models provide a credible simulation of the mean seasonal structure of the large-scale circulation, although a few models are outliers. (2) The models display a characteristic spread in their simulations that is nearly independent of latitude and season. (3) The seasonal differences among the models are greater than the variability among ensemble simulations of a single model. (4) Outliers aside, no single model performs best in all respects. (5) The reasons for a particular model's behavior are difficult to determine from the AMIP data alone, although there does not appear to be a strong dependence on resolution. (6) There are common seasonal errors apparent in some variables, although the observational data are of uncertain reliability in many cases. (7) While models on the whole appear to be slowly improving, the lack of systematic performance data prior to AMIP makes estimation of the rate of improvement difficult. Other parts of the standard output and the preliminary results from the diagnostic subprojects show that there are strong geographical variations in the accuracy of the models' performance, and that both the accuracy and agreement among the models generally decrease as smaller regions and shorter time scales are examined. It is also clear that a single statistical measure cannot adequately portray model performance. Any extension or expansion of AMIP should include improvements in the experimental design and a wider range of standardized output coordinated with reanalysis in order to permit more insightful diagnosis and accelerated model improvement.