Schematron V2 lands on OpenRouter: cheap, fast HTML-to-JSON extraction
Inference.net's closed extraction models now route through one gateway: 8B-class accuracy at 3B speed, 128K context, and a typed scraper for cents.
Copy markdown
Two tiers, one gateway away
Schematron V2 Turbo and Small went live on OpenRouter on Sept 12, so Inference.net's closed HTML-to-JSON models now sit behind the same gateway you already use for everything else. Turbo runs $0.03/M input and $0.15/M output; Small is $0.05/$0.23. Both carry a 128K-token context.
8B accuracy at 3B speed
On Inference's own extraction eval, V2 Small scores 4.06 (1-5 LLM-judge), matching the original 8B's 4.07 while running at 3B speeds. Turbo pushes 4.14 requests/sec on a single H100, roughly 2.5x the old 8B. Same JSON quality, far more pages per dollar.
Why not just point a general LLM at it
It's purpose-built for messy pages, so it skips the token bloat and latency of aiming GPT-6 or Claude at raw HTML. On SimpleQA (factuality inside web-search pipelines) Small hits 83.1. For price monitoring, filing parsing, or feeding RAG, this is the workhorse tier, not the frontier tax.
Build this weekend: a typed scraper for cents
Hand it a page plus a JSON schema and get clean, typed rows back. At $0.03/M input, a 10K-token page costs a fraction of a cent. Wire OpenRouter to a schema and a queue and you have a large-scale extraction pipeline with no model to host.