Results from the European Centre for Medium-Range Weather Forecasts (ECMWF) high-resolution (T213L31) global analysis-forecast system are analysed as a function of scale. Three regimes of forecast skill are identified, only one of which exhibits classical predictability behavior in which error, initially concentrated at smaller scales, penetrates up the spectrum and saturates at values roughly twice the observed variance. Two of the regimes are seen at spectral ranges previously accessible in global analyses and forecasts (roughly for spherical harmonic wavenumbers n < 80). The large-scale low-wavenumber regime (n < 10) is dominated by stationary (largely zonal) structures that are relatively uncontaminated by error up to the 10-day forecast limit. These structures represent “climatological” aspects of the flow, and their simulation is not considered to represent useful forecast skill. The intermediate wavenumber range (10 < n < 80) exhibits classical predictability behavior. A third regime at high wavenumbers (n > 100) unexpectedly, and in contrast to the classical predictability regime, exhibits forecast skill up to the 10-day forecast limit. The apparent enhancement of predictability at these small scales is due to local topographic forcing in the ECMWF analysis-forecast system.