best 9 document extraction apis

ibm released docling under mit with the pretrained layout and tableformer models under cdla-permissive-2.0. two separate licences, both genuinely permissive, and you should know you're accepting two rather than one. the project now sits under the lf ai & data foundation.

the technical report gives numbers rather than adjectives: 93.6% accuracy on table-structure recognition for tableformer, and roughly 95.4% on simple tables and 90.1% on complex ones for layout detection, evaluated against a doclaynet-derived corpus. most vendors in this ranking publish no self-benchmark at all.

being self-hosted answers the field that matters most here by construction — your contracts and medical records never reach a vendor, so there is no retention policy to read and no training clause to opt out of. what you give up is real: no sla, no support, no managed scaling, and throughput is your infrastructure's problem.

zero data retention is stated on every tier, including the free one. that is rarer than it should be: aws trains on your documents unless you opt out, and several vendors here address retention nowhere at all. the business plan adds soc 2 type 2, hipaa, gdpr and iso 27001 with dedicated infrastructure.

the pricing has a shape nothing else here does. $0.03 a page is mid-to-high, but the pay-as-you-go bill caps at $3,000 a month and anything beyond that up to a million pages is processed at no charge. at high volume the effective rate collapses toward zero, which inverts the usual penalty for scale.

the free allowance is 15,000 pages every month with no card required, and the core library is apache 2.0 and self-hostable if you'd rather run it. the gap is measurement: a flat per-page rate regardless of document type gives no signal about whether hard scans and dense tables are handled differently, and unstructured publishes no accuracy benchmark.

$1.25 per thousand pages for basic parsing is the lowest hosted rate in this ranking, and 10,000 free credits a month is a real allowance rather than a trial. llamaindex states it does not charge differently for scanned pages versus text pages, which is more than most vendors will commit to.

the tiering is where visibility stops. better table, chart and image extraction requires the advanced or agentic modes, and the credits-per-page multiplier for those is not published anywhere. since those are the modes anyone with complex documents will end up using, the headline rate describes the case you least need help with.

on data handling: files are cached encrypted for 48 hours by default and caching can be switched off entirely, with soc 2 type 2, hipaa and gdpr compliance. one oddity worth noting — the starter plan at $50 a month and the pro plan at $500 share the same $5,000 monthly pay-as-you-go overage cap despite very different bases.

flat per-page billing with no token component and no complexity tiering: $4 per thousand pages for markdown output, $5 for the document ai endpoint that returns structured fields. the batch api halves both, which for archive conversion is the number that matters.

the price change is worth stating plainly because it is invisible in older writing: ocr 3 cost $2 per thousand pages and ocr 4 replaced it around 23 june 2026 at $4. anything published before then — including most comparison articles still ranking on price — is quoting a rate that no longer exists.

two gaps. mistral publishes no definition of what counts as a page, so whether a scanned image bills like a text page is unstated, and there is no quantified accuracy benchmark on its own pages beyond claims of high accuracy across 40-plus languages. an account-level zero data retention option exists, though its applicability to the ocr endpoint specifically isn't confirmed.

reducto built and published rd-tablebench: a thousand hand-labelled complex table images including scans, handwriting and merged cells, released openly, on which it reports 90.2% average table similarity for its own pipeline. publishing the dataset alongside the score is the right way to make a claim like that, and nobody else here does it.

the pricing is credit-based and complexity-driven, which is honest about how the work actually varies and useless for forecasting. simple pages cost one credit, complex vlm-enhanced pages up to four, so a thousand pages lands anywhere between $15 and $60 depending on a mix you can't assess until it's processed. deep extract carries a 30-credit minimum per request.

15,000 free credits a month is generous, and reducto publishes a trust centre referencing zero retention, on-premise, air-gapped deployment, hipaa and soc 2 — though the specifics beyond that page's existence weren't confirmable.

at $5 per thousand pages on the growth plan, chunkr is the cheapest venture-backed hosted option in this ranking, with a segmentation model that identifies eleven or more element types per page including titles, tables, formulas and captions.

the self-hosting story needs reading carefully. the repository is dual-licensed: agpl-3.0 for open use, or a paid commercial licence that removes the copyleft obligation. but chunkr's own docs state the self-hosted build uses weaker 'community' ocr and segmentation models than the proprietary ones behind the cloud api. so the free path exists and does not give you the product you evaluated.

zero data retention is described as customisable and on by default, with soc 2 and hipaa compliance and files held temporarily in google cloud until a configurable expiry. the pricing itself we could not read: chunkr's page renders through javascript, so the rates above are secondhand and labelled as such.

for prebuilt models, microsoft states that input data and results are deleted within 24 hours and are not used to train its models, and custom models train only on data in your own blob storage inside your subscription. against aws's opt-out training default, that is the stronger position by a distance, and for invoices and identity documents it is the deciding one.

the prebuilt catalogue is broad — invoice, receipt, id, tax forms, health insurance card, contract — each priced as its own tier, with layout including table and structure extraction as a single tier rather than the separate surcharge aws applies.

the pricing is unreadable, for the third azure product this week. every figure on the public page is a '$-' placeholder pending a region and agreement selection, so the roughly $10 per thousand pages above comes from secondary sources rather than microsoft. the free allowance of 500 pages a month is confirmed.

the per-feature pricing is genuinely the clearest here: every capability has its own published rate rather than hiding behind a tier. plain text extraction at $1.50 per thousand pages is the cheapest hosted ocr in this ranking by a wide margin.

then the same pages cost $15 with table detection and $50 with forms — ten and thirty-three times the base rate, as separate api calls rather than add-ons. a document set you priced as plain ocr becomes a different budget entirely once you need the structure, and that multiplier is the single most consequential number on this page.

the data default is the reason it ranks eighth. aws's standard ai-service terms permit customer content to be used to develop and improve textract, and turning that off requires an opt-out policy set at the aws organizations level — not a checkbox, and not the default. for contracts, medical records or identity documents that is a decision someone should be making deliberately. the free tier also expires after three months rather than resetting monthly.

at a glance marker is apache 2.0 and free. the code is. the model weights, downloaded from the companion surya repository, ship under a modified ai pubs openrail-m licence that grants free use only for personal projects, research, or organisations under $5 million in trailing-twelve-month revenue and under $5 million in total lifetime equity or debt funding — and prohibits use outright, at any size, by anyone offering a product that competes with datalab.

self-hosting does not avoid this, because the restriction attaches to the weights rather than the distribution. a well-funded startup running marker locally on the strength of the apache badge is in breach and owes datalab a commercial licence, and no revenue figure exempts a competitor.

the product itself is good and honestly measured: datalab reports 76.0% for balanced mode and 66.6% for fast on olmocr-bench, an external benchmark rather than one of its own, and marker 2.0 shipped in july with the new mode structure. hosted pricing is $4 per thousand pages, or $6 for high accuracy, with add-ons at $0.30 each and $5 of free credits to start.