OpenRouter launches Extractions API for typed data from documents

A field-based approach to document extraction

The API lets a developer describe the fields needed from a document, then returns one typed value for each field. The request includes the document in state and a set of named fields, each with a type and optional instructions explaining what the field means.

OpenRouter says the API supports nine field types: text, number, date, time, datetime, list, duration, estimate and forecast. Values are returned as JSON types that can be stored without additional parsing. For example, a list can contain text, numbers or dates, while a duration can use a unit ranging from seconds to years.

The initial launch includes extraction models from Cloudflare, Inception and ScaleDown. Developers can access the API through the documentation or test it in OpenRouter’s Playground.

Evidence, confidence and uncertainty are part of the output

The example uses Cloudflare’s Clairic model to process an invoice. It extracts the invoice number INV-4417, a total of 3850, the contacts Dana Ruiz and Priya Shah, and a computed due date of 2026-11-28. The due date does not appear directly in the invoice. Clairic calculates it from the issue date, 2026-10-14, and the net 45 payment terms.

Text values can include spans identifying where the answer appeared in the source. Developers can use those spans to display supporting evidence next to each result. Single values also include confidence and p_null, the model’s estimate that the document contains no value for the requested field. A field becomes null when its p_null passes the configured threshold, while a separate guess shows what the model would otherwise have selected.

For the invoice’s estimated payment time, Clairic returns a distribution rather than a single certainty. The result has a median of 45 days, with p10 at 42 and p90 at 47, meaning the interval contains 80% of the mass described in the example.

Support for images and PDFs

Images are supplied through assets, referenced by a key in state, and use the same field definitions as text extraction. Image results may identify values as coming from the image and can include spell_confidence for text read from the picture.

In an example using a receipt image, Clairic returned a merchant value, a total of 29.04, a payment date of 2026-10-24 and tax of 3.79. OpenRouter says the API returns values rather than image coordinates, so the boxes shown in its demonstration are not part of the API output.

The company describes these confidence and evidence fields as a trust model. Developers can set a confidence floor, send lower-confidence results to a person and store the rest. The alpha release gives users a way to build document workflows around structured extraction while retaining signals about evidence, missing fields and uncertainty.