OpenRouter Adds Inference.net Schematron V2 HTML-to-JSON Extraction Models

OpenRouter quietly listed two new models from Inference.net in the early hours of September 12: inference-net/schematron-v2-turbo and inference-net/schematron-v2-small — a pair of 3B-parameter models built for exactly one job, HTML in, JSON out. No changelog entry; they surfaced in the public /api/v1/models response with created timestamps around 01:35 UTC.

What shipped

Two purpose-built HTML-to-JSON extraction models, 128K context, text-only. The interface is unusual: extraction instructions are supplied through a JSON schema in response_format, not a system prompt — structured_outputs and response_format are both in supported_parameters. Your prompt carries the page; the schema carries the instructions.

What changed — the pricing math

Turbo is $0.03 in / $0.15 out per million tokens with an 8,192-token output cap, tuned for high-volume runs. Small is $0.05 / $0.23, pitched at extraction quality for complex schemas and long pages. Compare the cheapest OpenAI generalist on the router, openai/gpt-5.6-luna at $0.20/$1.20 — Schematron Turbo is 6.7x cheaper on input and 8x on output. Against flagship gpt-6-astra ($10/$50), it's 333x. Worked example: extracting 1M pages at ~20K input / ~1K output tokens each costs roughly $750 on Turbo, $5,200 on Luna, and $250,000 on Astra. The caveat is quality — a 3B specialist will not match a frontier model on gnarly pages, which is exactly why Small exists.

Why a builder cares

Watch the Small output cap: it costs 53% more per output token than Turbo yet tops out at 4,096 tokens versus Turbo's 8,192 — a "long pages" variant that truncates JSON at 4K is a real trap for verbose product-page schemas. If you're paying frontier rates to turn scraped HTML into structured data, benchmark these on your worst 100 pages first; the price gap is large enough to fund a fallback pipeline entirely.