What document recognition actually does now
Text extraction alone is not enough when you need fields, tables and document layout. The value lies in the structure extracted from the document — fields, tables, paragraph boundaries, sometimes the coordinates of every fragment — precisely enough that an agent can cite a specific place instead of "reading the whole file".
That is why price in this market follows what you ask to come back, not how difficult the document is. Plain text, layout, and extracted fields are three different lines on a price list — and a difficult scan does not change the rate.
Three variants, different obligations
**A cloud API.** You pay per page and maintain nothing. The document leaves your network, the provider becomes a processor, and the documentation that goes with that becomes yours to keep.
**The vendor's containers on your own hardware.** Of the three largest clouds, only Microsoft publishes document-recognition containers. Your document does not travel, but the system still connects outward to report usage — cutting off the documents does not cut off the metadata. Nor is it a cheaper version of the cloud: Azure states that container pricing is the same as cloud service pricing, so you pay the list rate per page and add the hardware, the power and the people on your side. The models inside containers are updated less often than the hosted service.
**Open weights on your own GPU.** The weights are public, so neither the document nor the query has to leave the network — provided you also disable updates and telemetry yourself. The cost changes shape: instead of a rate per page you pay for the card's running time and for maintaining the setup.
These three differ by more than price. They differ in who is accountable for what came out of the document.
What we compare the cost against
Public vendor price lists in September 2026 started at $1.50 per thousand pages of plain text, with structure and tables around $10 per thousand pages. Those are list prices, they fall with volume, and they change every few months — so each one has to be checked with the vendor before it becomes a budget.
On the self-hosted side there is one published measurement: a hundred thousand pages processed on rented serverless infrastructure with open vision-language models, at $2.27 per thousand pages with a failure rate under two percent. That is a single run at full occupancy, and it describes an arrangement billed per use — it says nothing about the bill for a card you bought or for one kept rented by the hour, because those are two different billing models.
For plain text the reference point comes from a vendor's own arithmetic: AWS prices a hundred thousand pages at $150 in its worked example. That is different work from the measurement above, and comparing the two numbers would mislead.
Where the traps are
First: time, not documents. A rented GPU is billed by the hour, so you also pay for the gaps between batches. The bill then follows how evenly documents arrive, not how many pages there are.
Second: charges that do not appear on the first page of a price list. Keeping an idle deployed processor at a provider can be charged annually per version. Some services bill in ten-page blocks, and document-size limits on synchronous requests mean that above them, stitching pages is your problem rather than the service's.
Third: accuracy claims. In this market they are mostly self-reported; until a third party has measured a model, "99%" is a marketing sentence, not data.
When the cloud is fine
When the documents may go to a provider under a contract and you need ready-made field extraction, classification and a review interface, the cloud is a reasonable choice. Compare the full cost of doing the task yourself against the per-page price then — not price against price, because one of the two contains nobody's work and no hardware.
When the architecture is the answer
Where regulation or company policy does not allow a document to be sent outside the organisation — not merely where the industry sounds serious — no price list changes that condition. What remains is a setup running on hardware kept inside the company, along with responsibility for accuracy that you have to maintain yourself. A vendor container that connects outward to report usage has to be checked against that condition before it counts as satisfying it.
We do not sell a disconnected environment as the cheaper option. We treat it as a choice about control, with an operational price — and sometimes as the only choice a company is allowed to make.
Sources and the scope of the numbers
- Vendor price lists read on 28 September 2026: Azure AI Document Intelligence (https://azure.microsoft.com/en-us/pricing/details/ai-document-intelligence/) — text read $1.50 per thousand pages, layout models $10, extraction $30, container pricing equal to cloud pricing; AWS Textract (https://aws.amazon.com/textract/pricing/) — $1.50 per thousand pages to read, $65 for forms and tables; Mistral (https://mistral.ai/pricing/) — document recognition per thousand pages, with a batch discount. - The self-hosting measurement: "OCR'ing 100,000 pages with open-source VLMs on Modal", 2026 (https://www.redspring.xyz/blog/vlm-ocr-bench) — 100,000 pages, $223, under 2% failures, one run at full occupancy. - A calculator comparing self-hosting with APIs: https://gigagpu.com/ocr-cost-calculator — used as a point of reference, not as a source of prices.
This category re-prices often, and none of these figures is a quote for a particular deployment.
What this means for your company
Start from the documents, not from the tool. Split them into three piles: those that may go to any provider; those that may go only to a chosen one, under a contract; and those that cannot leave the company in any form. The second pile narrows the choice to providers willing to sign a processing agreement with a stated location; the third rules some options out by definition — and that pile is usually the one that settles the architecture.
Then check how many documents you process and whether they arrive steadily or in batches. If they arrive in batches, add the billable idle hours — those, not the rate per page, decide the bill.