HTML TO JSON3B PARAMETERSSCHEMA IS THE PROMPT128K CONTEXTCLOSED WEIGHTS
Schematron V2 Small API Pricing
Schematron V2 Small is Inference.net's own extraction model: send raw HTML and a JSON Schema, get back JSON that conforms to it - there is no prompt to write. It costs $0.05/M input and $0.23/M output, and Inference.net bills a cache read at the input rate, so caching saves nothing. It is the quality tier of the pair; Schematron V2 Turbo does the same job faster for less.
Input - per 1M tokens
$0.05/M
V2 Turbo $0.03/M the HTML
Output - per 1M tokens
$0.23/M
V2 Turbo $0.15/M the JSON
Cached input - per 1M tokens
$0.05/M
Same as input - no cache discount 0% off
Effective - agentic blend
$0.0644/M
92/8 split - cache hits cost the same
§ 01 / TERMINAL
Run the numbers.
Calculator pre-loaded with Schematron V2 Small's rates. There is no cache setting: Inference.net bills a cache read at the input rate, so caching changes nothing. In extraction the HTML going in is usually many times the JSON coming out, so the input rate carries most of the bill - and pre-cleaning the HTML is the cheapest saving there is.
$ /mo
Workload split
Tokens you can process
—
Words equivalent (English)
—
Effective rate
—
§ 02 / SCENARIOS
Real-world presets.
PRODUCT PAGE
Product page to JSON
$0.00074/page
LISTINGS
Results page to rows
$0.00219/page
ARTICLE
Article with metadata
$0.00118/page
LONG PAGE
Long page, one schema
$0.00496/page
Estimate · schematron-tokenizer-estimate · ≈3.75 chars/token Auto-counts as you type
This is a chars-per-token approximation, not a real tokenizer. Actual tokens vary by language, code density, and tool-call overhead — counts are typically ±10–20% off for English prose, more for code or non-Latin scripts. For exact billing, use the vendor's official tokenizer.
Characters 304
Words 55
Tokens (estimated) 81 tokens
Cost as input · uncached $0.000004 USD
Cost as output · uncached $0.00002 USD
| Model | Input /M | Output /M | Effective blended | Context | Best for |
|---|---|---|---|---|---|
| Schematron V2 Small Current | $0.05 cache $0.05 | $0.23 | $0.0644 agentic 92/8 | 128K | Best Schematron quality on complex schemas |
| Schematron V2 Turbo | $0.03 cache $0.03 | $0.15 | $0.0396 −39% | 128K | The same job at higher throughput |
| Ministral 3 3B | $0.10 cache $0.01 | $0.10 | $0.0321 −2.0× | 128K | A general 3B model at one flat rate |
| GPT-6 Luna | $0.10 cache $0.01 | $0.50 | $0.0682 +6% | 1.05M | OpenAI's small tier with structured outputs |
| Qwen3.8-Flash | $0.15 | $0.47 | $0.176 +2.7× | 1M | A general flash model with a 1M context |
| MiMo-V2.6-Flash | $0.14 cache $0.0028 | $0.28 | $0.0477 −26% | 1M | Cheap multimodal flash with deep cache discount |
Frequently asked.
What Schematron V2 Small costs, how it differs from Turbo, and what Inference.net does not price.
Q · 01 How much does Schematron V2 Small cost? +
$0.05/M input and $0.23/M output on Inference.net's model page, with cache reads at the input rate. On our 92/8 blend that is $0.0644/M; the scenarios above put that in pages per $100.
Q · 02 How do you call it? +
Through Inference.net's OpenAI-compatible API at
https://api.inference.net/v1. Send the page's HTML as the user message and your schema as a JSON Schema in response_format; the model does not read system or user prompts, so the schema's field names and descriptions are the whole instruction. Inference.net recommends stripping scripts, styles and boilerplate with lxml first, which also cuts the input tokens you pay for.Q · 03 Small or Turbo? +
Small is the quality tier: Inference.net scores it 4.060 on its own GPT-5.4-judged 1-5 extraction scale, against 4.039 for Turbo, and describes it as the one for complex schemas and long pages. Turbo is faster - 4.14 against 2.47 requests per second on one H100 in the vendor's test - allows 8K output tokens against Small's 4K, and costs $0.03/M input and $0.15/M output. Both scores are the vendor's own and have not been reproduced independently.
Q · 04 Why does prompt caching not lower the bill? +
Because Inference.net prints a cache-read rate equal to the input rate for both Schematron V2 models. A cached token costs the same as a fresh one, so the effective rate on this page is simply the blend of input and output, whatever cache-hit rate you assume. For extraction that rarely matters: every page is new HTML, so there is little to cache anyway.
Q · 05 Has the price changed since launch? +
Yes, downward, without an announcement we could find. The April 16, 2026 launch post priced it at $0.10 input and $0.25 output per 1M. The earliest Internet Archive capture of the model page, May 18, already reads $0.05 input, and a September 19 capture shows output at $0.23. The chart above plots each of those dated readings.
Q · 06 Are the weights open? +
No. The April 2026 launch post says V2 Small and V2 Turbo stay closed-source and are served only through Inference.net's API. The first generation, Schematron-3B and Schematron-8B, is open on Hugging Face (released August 2025), but those two are no longer on Inference.net's serverless model list, so there is no per-token price for them.
Q · 07 What are the rate limits, and is there a batch discount? +
Inference.net's API docs list 30 requests per minute on the free tier, 250 once you buy credits, and 1,000 on the Growth and Enterprise plans. The async endpoint (
/v1/slow) and the Batch API complete within 24 hours and are described as significantly cheaper, but no rate is published for either, so this page does not model a discount.