Apple's ToolSandbox benchmark reveals a significant performance gap between proprietary and open-source AI models, challenging recent claims and exposing weaknesses in real-world task execution.