In 2026, Anthropic engineer Thariq Shihipar argued that HTML is often a better format for AI-generated artifacts intended for human consumption, challenging the long-held preference for Markdown in working with large language models. Markdown became the default due to its lightweight syntax, readability, token efficiency, and seamless integration with version control systems like Git. One comparison estimated approximately 8,000 tokens for HTML versus 2,800 tokens for an equivalent Markdown version, while other studies reported token reductions of 20–30% when converting HTML into clean Markdown for LLM workflows. These savings lowered inference costs and freed context for reasoning, making Markdown popular for retrieval-augmented generation (RAG) pipelines. However, Shihipar’s argument, based on his experience building with Claude Code, was that HTML often provides a better experience when AI-generated artifacts are intended for people, not that Markdown had become obsolete.