Current models are quite far away from human-level physical reasoning (paper below). An upcoming version of models trained on world simulation will probably do much better.
PHYBench: Holistic Evaluation of Physical Perception and Reasoning in Large Language Models