CO2-induced changes in surface air temperature, precipitation, and sea level pressure are compared between model experiments participating in the second phase of the Coupled Model Intercomparison Project (CMIP2). A statistical formalism is applied, in which the average squared amplitude of the simulated climate changes is divided into a common signal and variances associated with internal variability and model differences. In the 20-yr period centered at the doubling of CO2 and for a set of 14-15 models, the dimensionless global relative agreement on gridbox-scale annual mean climate changes is 0.89 for surface air temperature but only 0.22 for precipitation and 0.46 for sea level pressure. A majority of the interexperiment differences are attributed to model differences; the contribution of internal variability to the differences in change is estimated as 16% for temperature, 34% for precipitation, and 32% for sea level pressure. For seasonal rather than annual climate changes, the agreement is lower and the contribution of internal variability to the interexperiment variance larger. Likewise, the relative agreement is worse and internal variability in relative terms more important earlier during the transient experiments than around the doubling of CO2. Conversely, when climate changes are averaged over larger areas than individual grid boxes, the relative agreement improves with increasing averaging domain (especially with precipitation and temperature) and the impact of internal variability decreases.