Flood levees are usually designed for a discharge of a given return period, whereas the uncertainty of the hydraulic parameters that convert discharge into water level is rarely included in the design. This study assessed the annual overtopping probability of hypothetical levees on three rivers with contrasting flood regimes: the regulated Platte River (Nebraska, USA; 51 years), the Choptank River (Maryland, USA; 32 years) and the Ngaruroro River (New Zealand; 31 years). Monthly maximum daily discharges were extracted from open daily records, and annual maxima were fitted with the generalized extreme value (GEV) distribution by L-moments. A seasonal model based on monthly GEV distributions was also tested. Each levee crest was set at the Manning depth of the nominal 100-year flood plus 0.5 m freeboard. The limit state included the annual maximum discharge, Manning roughness, bed slope, channel width, a model-error factor and the crest level. The failure probability was computed by the first- and second-order reliability methods (FORM, SORM), crude Monte Carlo simulation (MCS, 5×10⁶ samples) and importance sampling (IS, 2×10⁴ samples). The MCS failure probabilities were 4.45×10⁻³ to 8.16×10⁻³. FORM was within 2–15% of MCS, and the Hohenbichler–Rackwitz SORM within 0.7%, while needing only 92–279 limit-state evaluations. Hydraulic uncertainty raised the failure probability by 1.15–1.58 times at the heavy-tailed Platte and Choptank. At the Ngaruroro, whose fitted GEV is bounded above, overtopping was impossible with discharge as the only random variable, and roughness and model error contributed 48% of the FORM importance. Parametric bootstrap gave 90% intervals of the failure probability spanning more than an order of magnitude. A significant upward trend in the Choptank annual maxima (p = 0.012) tripled the failure probability between 1980 and 2011 under a non-stationary GEV. The seasonal model overestimated risk because monthly maxima were serially correlated. FORM and SORM are efficient and accurate for levee overtopping, but parameter uncertainty, the choice of the distribution and non-stationarity affect the failure probability more than the choice of the reliability method.