Expand Fetch claims 74% vs 55% on hard web QA, giving agents vision-based page reading
rohanpaul_ai · x · 2026-10-15
AI startup Expand launched Expand Fetch, a web scraping/crawling API built for AI agents that aims to fix the "agents never really see the web" problem.
- The problem: conventional pipelines flatten pages to plain-text Markdown, dropping icons (like a ✗ next to a feature), tables, and hidden data, so agents confidently answer wrong.
- Two in-house models: Iris, a vision model that reads pages like a human and preserves tables, icons, and layout; X-Ray, a JSON model that uncovers hidden data behind pages.
- Traceable snapshots: every fetch is stored with Markdown lines linked back to source, so developers can trace a wrong answer to exactly what the agent saw.
- Benchmarks: on the company's MDBench Hard (1,094 questions over 326 challenging pages, answered by Claude Sonnet 5.5, strictly graded), Expand scored 74% vs 55% for the next best API Exa and 27% for Firecrawl. Iris hit 92.72% average F1 on the validation set versus 69.55% for comparison model Jev.
Related event: Expand Launches Expand Fetch to Let AI Agents Truly 'See' Web Pages(4 posts)→
More from coding & agent
- Dev bills $20,000 for an app built with $200 of Claude Code — IgorCarron · 2026-10-15
- Dan Shipper: run product and labs in parallel, even with a labs team of one — danshipper · 2026-10-15
- Gemini 4 Argon wrote 5,000+ lines of pure Python to build its own ray-tracing 3D engine — aitrendz_xyz · 2026-10-15
- Minecraft fully ported from Java to Rust, file by file, server and client included — Vjeux · 2026-10-15
- mitsuhiko builds a full Python IDE for iOS in an afternoon with new AI models — mitsuhiko · 2026-10-15
- How to get Gemini 4 Argon to build games that actually look decent — chongdashu · 2026-10-15