The document introduces a new game environment called Mazebase that is designed as a sandbox for machine learning approaches to reasoning and planning. Within Mazebase, the authors create 10 simple 2D maze games based on algorithmic tasks like following if-then statements. They deploy neural models on these games to evaluate how well the models can learn the tasks. The best performing models can learn to complete some non-trivial tasks in Mazebase, but still fail to learn all of the games' aspects. The controlled nature of Mazebase allows analyzing where exactly models succeed and fail.