Experimental tests of choice predictions in one-shot games show only little support for Nash equilibrium (NE). Poisson Cognitive Hierarchy (PCH) and level-k (LK) are behavioral models of the thinking-steps variety where subjects differ in the number of levels of iterated reasoning they perform. Camerer et al. (2004) claim that substituting the Poisson parameter τ=1.5 yields a parameter-free PCH model (pfPCH) which predicts experimental data considerably better than NE. We design a new multi-person game, the Minimizer Game, as a testbed to compare initial choice predictions of NE, pfPCH and LK. Data obtained from two large-scale online experiments strongly reject NE and LK, but are well in line with the point-prediction of pfPCH.