Result: A Comparative Study of Vibe Coding with ChatGPT and Gemini in Front-end Web Development
Further Information
The invention of Generative AI and Large Language Models has recently catalyzed “vibe coding” as a new paradigm of software development in which developers use natural language to state their intentions. However, there is currently a significant lack of empirical research comparing the fundamental behaviors of GenAI tools and their code quality. This paper presents such comparative study of GPT-4o and Gemini 2.5 Pro for front-end web development using everyday technologies HTML, CSS, and JavaScript. Using zero-shot and prompt-chaining strategies, we tasked the models to create three commonplace web applications of increasing complexity. The architecture and features of the generated code were evaluated using a mixed-method evaluation framework. The results show that GPT-4o and Gemini 2.5 Pro represent two different development paradigms; GPT-4o functions as a tool that generates minimal, concise code that follows user instructions, but in more complex tasks it sometimes generates errors and unwanted changes in the codebase. In contrast, Gemini 2.5 Pro operates as a proactive-enhancement agent that generates more complex, feature-rich code by anticipating the user's needs and adding advanced UI functionalities. Importantly, vibe coding is formally defined, explained, and compared to other AI-assisted programming approaches. The codebase created for this research is available at: https://github.com/mhorvat/vibecoding_frontend.