Autonomous navigation in unstructured environments requires not only efficient path planning but also a robust understanding of terrain conditions. In this paper, we propose a novel framework that combines the classic A* algorithm with a vision-based traversability analysis pipeline powered by a visual-language model (GPT-4o). The robot periodically captures images of its environment and transmits them to a server, where GPT-4o evaluates terrain types and generates semantic traversability scores. These scores are dynamically incorporated into the A* search space as cost modifiers, allowing the robot to avoid risky or inefficient routes based on real-time visual input. We validate our framework in a simulated Unity environment and demonstrate its ability to adapt navigation decisions based on terrain complexity, enabling safer and more intelligent autonomous driving.