Claude Leads LLM Web Generation in 8-Week Study

An article recently published on arXiv (2605.06707) details an observational study spanning eight weeks, comparing 68 single-file HTML outputs from four families of reasoning models—GPT, Gemini, Grok, and Claude—gathered during 17 public experiments in the "HTML AI Battle" initiative from December 10, 2025, to February 4, 2026. Each output was assessed based on rendered browser video, utilizing human evaluations and a Gemini LLM-as-a-judge layer to measure prompt adherence, functional accuracy, and UI quality, all conducted under a standardized public-interface protocol without custom instructions or personality adjustments. The results were formatted for social media platforms, including X (Twitter), TikTok, and YouTube. Claude emerged as the top-performing family, consistently achieving the highest mean scores and winning the most comparisons.