The short answer
Select the response shape and operating limits first. Measure accuracy, latency and cost on the same representative document set before routing results into business actions.
An OCR API can return recognized text, a layout model, invoice fields or a response shaped to your own schema. Those outputs are not interchangeable. Choose the contract your application needs before comparing page rates, then check how multi-page files, asynchronous requests and uncertain fields move through your system.
Textract, Azure and Google offer cloud-native building blocks with distinct models and request rules. Mindee and Dokyumi emphasize schema-shaped extraction; ABBYY adds a broader document-skill platform and deployment choices. This order is a contextual engineering shortlist, not a benchmark of recognition accuracy or latency.
All entries are based on official documentation and pricing, with no shared hands-on test. A production decision also needs your own permissions, storage, region, failure handling and review design. Raw extraction charges cover only part of that operating cost.
The shortlist at a glance
Prices use the billing basis shown. Page, task, credit and execution allowances measure different things. On a small screen, scroll the table horizontally.
| Tool | Best fit | Price and billing basis |
|---|---|---|
| 1. Amazon Textract | AWS-native text, forms and invoices | US West Oregon examples: OCR US$0.0015/page, Tables US$0.015/page, Forms US$0.05/page in the first million/month. Selected feature charges combine. |
| 2. Azure Document Intelligence | Microsoft-centered custom models | F0 includes 500 pages/month, but only the first two pages per request. Confirm model/region paid rates in Azure’s calculator. |
| 3. Google Cloud Document AI | Processor-specific document pipelines | Form Parser/Custom Extractor: US$30 per 1,000 pages in the first million/month. Invoice Parser: US$0.10 per document block of 1–10 pages. |
| 4. Mindee | Schema-led extraction in an application | Subscription credits: one per page, or 1.5 with confidence scoring. Default excess usage adds a 20% surcharge; unused credits expire. |
| 5. DokyumiSoxoa-owned product | Custom business-document JSON and portals | 25 credits/month free. Starter: US$99/month for 500 credits. One file uses one credit per five pages, rounded up; no automatic overages. |
| 6. ABBYY Vantage | Enterprise document skills and deployment | No standardized amount verified on the checked product page; request a scope-specific demonstration and quote. |
What matters for this decision
- Response contract
- Distinguish plain text, layout coordinates, pretrained fields and custom schemas; verify tables and confidence at the required level.
- Request lifecycle
- Check file size, page limits, online versus batch processing and how failed or partial results are identified.
- Model and deployment fit
- Evaluate training inputs, languages, regions and deployment options for the actual document class and organization.
- Billing semantics
- Normalize the intended features, page counts and model charges; do not compare a raw OCR rate with a full extraction workflow.
The ranked picks
Amazon Textract
Best for aWS-native text, forms and invoices
Amazon Textract exposes separate APIs for OCR, document structure and specialized invoice or receipt extraction. AnalyzeExpense separates summary fields from line-item groups; AnalyzeDocument offers forms, tables and queries. Developers select the features they need, then build the storage, validation and downstream workflow around the returned data.
Separate OCR, AnalyzeDocument and AnalyzeExpense APIs let an AWS team choose text, structural features or invoice fields rather than buy one undifferentiated extraction endpoint.
The tradeoff: Feature charges combine, and multipage PDF/TIFF needs the asynchronous route.
Price and billing: US West Oregon examples: OCR US$0.0015/page, Tables US$0.015/page, Forms US$0.05/page in the first million/month. Selected feature charges combine.
Check before choosing: Synchronous PDF/TIFF is single-page; multipage files require the asynchronous workflow.
Azure Document Intelligence
Best for microsoft-centered custom models
Azure Document Intelligence offers Read/Layout, prebuilt document models and custom extraction. Its invoice model returns billing fields and line items, while custom template or neural models use labeled examples. This fits teams integrating document data into Microsoft infrastructure who can manage training inputs, model versions and application-side review.
Read/Layout, invoice models and custom extraction fit teams that can prepare labeled examples and integrate the result into their existing Azure application lifecycle.
The tradeoff: F0 processes only two pages per request; confirm production model and tier limits.
Price and billing: F0 includes 500 pages/month, but only the first two pages per request. Confirm model/region paid rates in Azure’s calculator.
Check before choosing: Custom models need prepared examples; file limits and page processing differ by service tier.
Google Cloud Document AI
Best for processor-specific document pipelines
Google Document AI separates document processing into purpose-specific processors. Invoice Parser extracts invoice fields and line items; Form Parser returns generic fields and tables; Custom Extractor supports a defined schema. This gives an engineering team model choices, but those choices also change configuration, limitations and the billing unit.
Form, invoice and custom extractors provide distinct choices when the application needs to match each document class to an appropriate processor and schema.
The tradeoff: Processor restrictions differ; invoice billing blocks are not a universal per-page price.
Price and billing: Form Parser/Custom Extractor: US$30 per 1,000 pages in the first million/month. Invoice Parser: US$0.10 per document block of 1–10 pages.
Check before choosing: Form Parser cannot be up-trained and targets simple tables; processor/version limits differ.
Mindee
Best for schema-led extraction in an application
Mindee provides extraction APIs built around editable data schemas for catalog or custom document types. Schemas support typed fields, objects and arrays, allowing developers to shape the output for their application. Optional confidence features add review signals, while the surrounding intake, approval and business workflow remains an integration decision.
Typed fields, objects and arrays support a developer-defined output contract, with catalog schemas that can be modified for the business case.
The tradeoff: Check nesting, field guidance and the extra credit cost for confidence scoring.
Price and billing: Subscription credits: one per page, or 1.5 with confidence scoring. Default excess usage adds a 20% surcharge; unused credits expire.
Check before choosing: Schema nesting is limited; documentation recommends no more than 25 fields for latency.
Dokyumi
Best for custom business-document JSON and portals · Soxoa-owned
Dokyumi returns JSON shaped by a saved custom schema, with model-reported field confidence and validation details. Intake can use its dashboard, REST API or branded customer upload portals. The application receiving the result keeps control of business checks and approval; a processing status does not approve a transaction.
A saved schema and shared extraction endpoint fit applications that need defined business fields and optional customer upload portals alongside structured responses.
The tradeoff: It is higher-level extraction rather than generic OCR; Soxoa owns this product.
Price and billing: 25 credits/month free. Starter: US$99/month for 500 credits. One file uses one credit per five pages, rounded up; no automatic overages.
Check before choosing: Self-service files stop at 50 pages; confidence still needs document-specific evaluation.
ABBYY Vantage
Best for enterprise document skills and deployment
ABBYY Vantage organizes intelligent document processing around pre-trained and custom document skills. Its low-code designer, human review and connectors support workflows that feed enterprise automation systems. Cloud, on-premises and private Azure deployment options make it relevant where deployment control and existing RPA infrastructure matter alongside field extraction.
Its REST integration, reusable skills and cloud or private deployment options are relevant when extraction belongs in a larger governed document-automation platform.
The tradeoff: Scope skill configuration, operational ownership and commercial terms before comparing request prices.
Price and billing: No standardized amount verified on the checked product page; request a scope-specific demonstration and quote.
Check before choosing: Skill configuration and system integration make this a platform implementation, not a simple converter.
Choose by the constraint you have
- The application already uses one major cloud: Start with that cloud’s relevant modelCompare the integration and region fit before introducing another processing service.
- You need business fields in your own schema: Mindee, Dokyumi or a cloud custom extractorEvaluate schema depth, missing fields and review behavior on identical documents.
- Deployment control is a deciding requirement: ABBYY Vantage and documented cloud optionsVerify the exact supported deployment, residency and contract rather than inferring it from a security badge.
Build a reproducible API evaluation
- Freeze a permitted test set with known field values, representative languages, multi-page files and poor-but-realistic scans.
- Record model/version, features and request mode. Compare required-field correctness, false values, unresolved fields and complete-document success.
- Measure end-to-end latency and actual charges, including retries and surrounding storage or workflow services; preserve raw outputs for review.
- Exercise page-limit errors and unavailable destinations. Require explicit review rules before the application acts on uncertain extraction.
Keep the trial result and the original input together. A vendor feature list establishes what to investigate; it does not establish your accuracy, savings or operating cost.
Questions before you commit
Which API has the best OCR accuracy?
This guide does not establish a universal winner. Recognition and extraction quality depend on document class, layout, language, scan quality and selected model. Use the same representative set and separate missing fields from confidently wrong values when reporting results.
Is a low per-page rate the cheapest option?
Not necessarily. Forms, tables, custom extraction, confidence, hosting or workflow operations can have separate charges. Include retries and engineering operations, then compare the cost of a usable reviewed record rather than only the first request.
Can confidence scores approve the result automatically?
A confidence score can help route review, but it does not establish document authenticity or business authorization. Check how the provider defines the score and validate thresholds against labeled files before using them to control a downstream action.
Sources and research limits
We checked the official pages linked below on October 27, 2026. Recommendations combine those documented capabilities with the specific workflow in this guide. Features, rates and availability can change. Confirm the current terms and run the evaluation above before choosing.
Published prices are in US dollars unless explicitly stated. An annual-billing monthly equivalent requires the vendor's annual commitment. A quote-based plan is not treated as free. Missing or ambiguous prices are identified instead of estimated.
- Amazon Textract. AnalyzeExpense response · Document/API limits · Regional feature pricing. Checked October 27, 2026.
- Azure Document Intelligence. Models and input requirements · Custom-model training · Regional plans. Checked October 27, 2026.
- Google Cloud Document AI. Processors and supported fields · Form Parser limitations · Processor-specific pricing. Checked October 27, 2026.
- Mindee. Data schemas and constraints · Extraction models · Credit and overage terms. Checked October 27, 2026.
- Dokyumi. Workflow and output contract · Plans, credits and limits. Checked October 27, 2026.
- ABBYY Vantage. Skills, review and deployment. Checked October 27, 2026.