The adequate simulation of internal climate variability is key for our understanding of climate as it underpins efforts to attribute historical events, predict on seasonal and decadal timescales, and isolate the effects of climate change. Here the skill of models in reproducing observed modes of climate variability is assessed, both across and within the CMIP Versions 3, 5, and 6 archives, in order to document model capabilities, progress across ensembles, and persisting biases. A focus is given to the well-observed tropical and extratropical modes that exhibit small intrinsic variability relative to model structural uncertainty. These include the El Niño / Southern Oscillation (ENSO), the Pacific Decadal Oscillation (PDO), the North Atlantic Oscillation (NAO), and the Northern (NAM) and Southern (SAM) Annular Modes. Significant improvements are identified in models’ representation of many modes.Canonical biases, which involve both amplitudes and patterns, are generally reduced across model generations. For example, biases in ENSO-related equatorial Pacific sea surface temperature, which extend too far westward, and associated atmospheric teleconnections, which are too weak, are reduced. Stronger tropical expression of the PDO in successive CMIP generations has characterized their improvement, with some CMIP6 models generating patterns that lie within the range of observed estimates. For the NAO, NAM, and SAM, pattern correlations with observations are generally higher than for other modes and slight improvements are identified across successive model generations. For ENSO and PDO spectra and extratropical modes, changes are small compared to internal variability, precluding definitive statements regarding improvement.