This study analyzes the effect of increasing horizontal resolution in the atmospheric model ECHAM6 on the simulated mean climate state and climate variability. For that purpose three AMIP-style simulations with the resolutions T63L95, T127L95, and T255L95 are compared to reanalysis data and observations. Biases in atmospheric fields as well as tropospheric and stratospheric biases individually are analyzed. Besides mean errors of the climate state and the variance, some atmospheric phenomena with different time scales are studied at the three horizontal resolutions: the transient eddy kinetic energy, storm tracks, atmospheric teleconnections, the Madden–Julian-Oscillation (MJO), and the Quasi-Biennial Oscillation (QBO). The main result is that, overall, the bias of the simulated climate is reduced with increasing resolution when considering the mean state and the variance. A greater improvement takes place in the extra-tropical than in the tropical troposphere. The errors in the stratosphere are generally larger but the relative benefit of increasing resolution is greater than in the troposphere and we find that stratospheric phenomena, like the QBO, are sensitive to horizontal resolution. Globally, the bias of the mean state improves by 19 %, while the bias of the variability improves by 15 % (from T63 to T255). Major challenges remain the simulation of the precipitation and climate features like the MJO, which might require a coupled atmosphere–ocean model.