The main purpose of this paper is to investigate the ultimate behavior of steel cable-stayed bridges with design variables and compare the validity and applicability of computational methods for evaluating ultimate load capacity of cable-stayed bridges. The methods considered in this paper are elastic buckling analysis, inelastic buckling analysis and nonlinear elasto-plastic analysis. Elastic buckling analysis uses a numerical eigenvalue calculation without considering geometric nonlinearities of cable-stayed bridges and the inelastic material behavior of main components. Inelastic buckling analysis uses an iterative eigenvalue calculation to consider inelastic material behavior, but cannot consider geometric nonlinearities of cable-stayed bridges. The tangent modulus concept with the column strength curve prescribed in AASHTO LRFD is used to consider inelastic buckling behavior. Detailed procedures of inelastic buckling analysis are presented and corresponding computer codes were developed. In contrast, nonlinear elasto-plastic analysis uses an incremental-iterative method and can consider both geometric nonlinearities and inelastic material behavior of a cable-stayed bridge. Proprietary software ABAQUS are used and user-subroutines are newly written to update equivalent modulus of cables to consider geometric nonlinearity due to cable sags at each increment step. Ultimate load capacities with the three analyses are evaluated for numerical models of cable-stayed bridges that have center spans of 600 m, 900 m and 1200 m with different girder depths and live load cases. The results show that inelastic buckling analysis is an effective approximation method, as a simple and fast alternative, to obtain ultimate load capacity of long span cable-stayed bridges, whereas elastic buckling analysis greatly overestimates the overall stability of cable-stayed bridges.