SOM Outperforms HTML and Markdown in Web Agent Efficiency Tests

As AI agents increasingly interact with web pages, choosing the right page representation format becomes critical for both cost and functionality. Testing on GPT-4o and Claude Sonnet 4, researchers found that while Markdown offers the lowest token count, it sacrifices interactivity—making SOM (Semantic Object Model) the optimal choice for agents needing both efficiency and the ability to click, navigate, and interact with dynamic content.