We investigate the effect of solvent models on the computed thermodynamics of protein folding. Atomistic folding simulations of a fast-folding mini-protein, CLN025, were employed to compare two commonly used explicit solvent water models, TIP3P and TIP4P/Ew, and one implicit solvent (AMBER generalized Born) model. Although all three solvent models correctly identify the same native folded state (RMSD = 1.5 ± 0.1 Å relative to the experimental structure), the corresponding free energy landscapes vary drastically between water models: almost an order-of-magnitude difference is seen in the predicted fraction of the unfolded state between the two explicit solvent models, with even larger differences between the implicit and the explicit models. Quantitative arguments are presented for why the sensitivity is expected to hold for other proteins, as well as for other conformational transitions involving large changes in solvent exposed areas such as protein-ligand binding. Comparing protein-solvent and solvent-solvent contributions to the folding energy between different water models, water-water electrostatic interactions are identified as the largest contributor to the differences in the predicted folding energy, which helps explain the strong sensitivity of the folding landscape to subtle details of the water model. For the two explicit solvent models, differences in water model parameters also result in the average number of water molecules surrounding the protein being noticeably different. Water models that poorly reproduce certain bulk properties of liquid water such as self-diffusion are likely to misrepresent water-water interactions; we argue that within a pairwise additive energy function this error cannot, in general, be compensated by an adjustment to the solute-solute and solute-solvent parts of the energy.