Spur gears are one of the most commonly employed machine elements for power transmission. During their operating life, they are subjected to variable and cyclic loading. Consequently, material fatigue may occur in the tooth root region, more commonly referred to as bending fatigue. Steel gears are typically subjected to heat treatment processes such as carburizing to reduce wear. Additionally, the gears may be shot peened to induce a beneficial, fatigue-resistant surface layer of compressive residual stresses. However, this may result in bending fatigue crack initiation below the surface, i.e., subsurface bending failure. This type of crack is hardly detectable during regular service intervals. Hence, it may often go unnoticed, grow, and propagate rapidly through the brittle carburized layer, resulting in tooth breakage and gear failure. This doctoral thesis establishes a computational model for predicting the bending fatigue crack location (surface vs. subsurface) and the required number of cycles for surface-hardened spur gears. Due to the relative complexity of the gear’s geometry and loading conditions, the finite element method is employed to obtain load-induced stresses and strains. The multilayer method is employed to account for inhomogeneous material, while the strain life approach (ε – N) is used to predict the bending fatigue life. The research is divided into multiple stages. During the initial stages, bending fatigue with surface/subsurface crack initiation is investigated on relatively simple geometry, such as surface-hardened gear steel specimens. Then, the model is gradually upgraded to account for more complex loading conditions, residual stress distributions, and inhomogeneous material until the final model applicable to the running gear pair is acquired. According to the obtained numerical results and confirmed by the experimental investigations from the available literature, subsurface bending fatigue crack initiation tends to occur only in carburized and additionally shot-peened gears. The critical region for subsurface failure is located in the proximity of sharp loss of compressive residual stresses and still relatively high load-induced stresses. The probability of subsurface crack initiation can be reduced by modifying the gear’s geometry (choosing a smaller normal module) or increasing the beneficial compressive residual stresses by prolonging the carburization time.