We establish precise upper-tail asymptotics and large deviation principles for the rightmost eigenvalue $λ_1$ of Wigner matrices with sub-Gaussian entries. In contrast to the case of heavier tails, where deviations of $λ_1$ are due to the appearance of a few large entries, and the sharp sub-Gaussian case that is governed by the collective deviation of entries in a delocalized rank-one pattern, we show that the general sub-Gaussian case is determined by a mixture of localized and delocalized effects. Our key result is a finite-$N$ approximation for the upper tail of $λ_1$ by an optimization problem involving \emph{restricted annealed free energies} for a spherical spin glass model. This new type of argument allows us to derive full large deviation principles when the log-Laplace transform of the entries' distribution $μ$ has bounded second derivative, whereas previous results required much more restrictive assumptions, namely sharp sub-Gaussianity and symmetry, or only covered certain ranges of deviations. We show that the sharp sub-Gaussian condition characterizes measures $μ$ for which the rate function coincides with that of the Gaussian Orthogonal Ensemble (GOE). When $μ$ is not sharp sub-Gaussian, at a certain distance from the bulk of the spectrum there is a transition from the GOE rate function to a non-universal rate function depending on $μ$, and this transition coincides with the onset of a localization phenomenon for the associated eigenvector.