Modern energy grid expansion planning, by necessity, includes timeseries data to accurately model storage and renewable assets. Representative time periods are commonly used as a way to decrease problem size and therefore mitigate the increased complexity from this inclusion. However, there are many choices around these representative periods: length; location in planning horizon; boundary conditions. Each of these can influence what is considered the optimal planning outcome, potentially causing disparity with the original problem’s optimal solution. As an alternative to this approach, we present a novel methodology to contend with the entire planning horizon including the embedded timeseries by augmenting Benders decomposition (BD). To improve the speed of generating Benders cuts, the full subproblem can be partitioned into smaller independent subproblems covering the entire horizon. However, this approach ignores the boundary conditions between subproblems, hence they only produce a lower bound for the GSTEP problem. Here we define a multi-fidelity BD approach that converges quickly by using small subproblems for approximate objective evaluation and fast cut generation. However, it still guarantees optimality by ultimately solving a few full subproblems. We apply this approach to a variety of modified IEEE test cases and several scenarios of the real-world Australian National Energy Market to show that dynamically adjusting the subproblems’ size between multiple fidelities can provide significant speed-up. Although the proposed method is tested in the context of this specific case study, it is generically applicable to other long-term planning problems that use Benders decomposition.