Purpose: This study compares the performance of GPT-4o, Claude Sonnet 4, and DeepSeek V3 in front-end web code generation using a dynamic code testing framework, focusing on execution speed, code efficiency, functional performance, and output quality. Design/methods/approach: A mixed-methods comparative design was applied across five progressively complex front-end development tasks implemented using HTML5, CSS3, and JavaScript in a zero-shot setting. Quantitative evaluation covered execution time, code length, and Google Lighthouse scores, while qualitative assessment examined functional correctness, readability, semantic accuracy, interactivity, visual quality, standards compliance, and code efficiency. Three composite metrics, Efficiency Index (EI), Normalized Speed (NS), and Composite Dynamic Score (CDS), were used to integrate the evaluation dimensions. Findings: GPT-4o achieved the highest speed and efficiency, producing compact code with a mean completion time of 13.16 seconds and the highest CDS. Claude Sonnet 4 demonstrated stronger visual quality and interface design despite longer code and slower generation. DeepSeek V3 produced stable and functionally consistent outputs across task levels but recorded comparatively slower execution. The results reveal clear trade-offs between speed, efficiency, and qualitative output characteristics. Research implications/limitations: The findings provide practical guidance for selecting AI models according to development priorities, including rapid prototyping, interface quality, and output consistency. However, the single-run experimental design, with one output per model for each task, limits statistical generalizability. Future studies should employ repeated trials, broader task sets, and additional models. Originality/value: This study proposes an execution-based, multidimensional framework that combines dynamic testing, qualitative assessment, and composite performance metrics, offering a structured approach for evaluating AI-generated front-end code beyond conventional static assessment.
Copyrights © 2026