Document extraction: hosted API or open model?

Reading text and structure out of scanned documents used to mean either a paid cloud API or a lot of custom work. That gap has narrowed. In October, LightOn released LightOnOCR-3, a family of open-weight document models under the Apache 2.0 licence, following Mistral's release of OCR 4 as a paid API in June. For a team automating invoices, forms or contracts, choosing between a hosted service and a model you run yourself is now a real decision, and the right answer depends on your documents, your volumes and where your data is allowed to go.

What changed this year

Mistral OCR 4, announced on 23 June 2026, returns a bounding box for each block of text, classifies blocks such as titles, tables and signatures, and gives confidence scores per page and per word. Its announcement lists API pricing of $4 per 1,000 pages, or $2 through the batch API, and $5 per 1,000 pages for the Document AI layer, which returns JSON matching a schema you define. Self-managed deployment is offered to enterprise customers.

LightOnOCR-3, announced on 8 October 2026, comes in 0.8-billion and 4-billion-parameter versions that LightOn says are available for commercial use, modification and redistribution. The model card describes two modes: plain transcription to markdown, and a grounding mode that labels each block and gives its position on the page. LightOn's press release says the model ranks second on olmOCR-Bench behind a model more than eight times its size, and that self-hosted processing can cost below one cent per thousand pages, depending on hardware, usage and workload.

Treat benchmark scores as a starting point

Both vendors publish strong benchmark results, and both sets of results come from the vendors. Mistral's announcement is candid about this: it says public document benchmarks contain scoring artefacts, such as errors in the reference answers and reading-order problems in multi-column layouts, describes the scores as directional and recommends testing on your own documents. That advice applies to either option.

A public benchmark mixes academic papers, forms, tables and handwriting. Your workload is usually narrower: supplier invoices in a dozen layouts, say, or delivery notes photographed on a phone. A model that wins overall can still miss the one field you care about, such as a VAT number in a footer. Build a test set from 50 to 100 of your own documents, including the awkward ones, write down the correct value for each field you need, and score every candidate against that.

Where the costs really sit

Per-page prices are easy to compare and rarely decide the matter. At Mistral's listed rate, 10,000 pages a month costs $40 through the standard API, or $20 through batch. Running an open model yourself removes that charge but adds a server with a suitable GPU, someone to keep it updated and monitored, and the time to connect it to your systems. LightOn's below-one-cent figure covers processing only, not that work.

For small and medium volumes, the model is often the cheap part. The larger costs tend to be checking outputs, handling exceptions and fixing the process when a supplier changes their template. Plan for those whichever option you choose.

Data handling can matter more than accuracy

If your documents contain personal data, contracts or financial records, where they are processed may matter more than a few points on a benchmark. With a hosted API you send each page to a third party, so you need to know where it is processed, how long it is kept, whether it is used for training, and what your contract says. With an open model on your own infrastructure, the pages need not leave your environment, which can simplify your data protection assessment, although you then own the security of that server.

Open models have their own constraints. The LightOnOCR-3 model card says the model supports only its two built-in modes, that other instructions are outside what it was trained on, and that PDFs should be rendered to images at a recommended resolution first. Read those notes before you plan a pipeline around any model.

A short checklist before you choose

• Collect 50 to 100 real documents, including poor scans and unusual layouts, with the correct value for each field you need.
• Score each candidate on those fields, not on overall text accuracy.
• Use confidence scores and simple validation rules, such as totals that add up and dates that parse, to send uncertain results to a person.
• Estimate monthly volume and compare API charges with the full cost of running a server, including staff time.
• Check where documents may be processed and what your contracts and privacy notices allow.
• Keep the extraction step behind a simple interface so you can change models later.

Document extraction is a practical place to start with AI, because each result can be checked against what a person would have typed. If you are weighing a hosted service against a model you run yourself, we can help you test both on your own documents and design the workflow around the one that fits. See our NLP and document workflows and Own Your Stack services, or get in touch.