The whole token tax hits 7.4x median thing and straight-up admitting that AI can't actually fix blocking issues is the stuff vendors conveniently gloss over in their shiny demos. But the real magic in production isn't picking one approach and riding with it - it's stacking them smart like API - deterministic selector - AI fallback - vision. The McGill benchmark definitely backs up your accuracy numbers, tho I'd eyeball that traditional scraper baseline a bit closer if I were you. Your maintenance cost breakdown 10-15% weekly is legit, but it swings pretty hard depending on how volatile your target sources are, so people should really run the numbers on their own stuff. And honestly that self-healing selector pattern deserves way more spotlight , regenerate on null, validate with Pydantic, like $0,0003 per heal -that's literally why layered systems are crushing it in production rn.
Datablue
Turn Any Website Into LLM-Ready Data
The distinction between fetching and parsing is an important one. AI extraction can make the parsing layer much more flexible, but it still depends on reliably retrieving the page first. In production, having a clean separation between crawling/fetching, content extraction, and structured output makes the overall pipeline much easier to maintain.