7 Commercial Players

Executive summary

The commercial landscape divides into four tiers in 2026: AI-native specialists (LlamaParse, Reducto, Landing AI ADE, Mistral Document AI), hyperscaler APIs (Google Document AI, AWS Textract, Azure Document Intelligence), classic IDP retrofitted (ABBYY, Hyperscience, UiPath, Klippa, Rossum) and AI-native commercial IDP (Nanonets, Docsumo), and the newer platform-integrated tier (Databricks Document Intelligence).
AI-native specialists lead on benchmark scores and pricing flexibility; hyperscaler APIs lead on procurement maturity and broad compliance certification; classic IDP leads on HITL and audit posture; the platform tier is a new strategic shape that makes the shift-left thesis institutionally credible.
The right vendor is rarely the one with the highest benchmark score. The right vendor is the one that minimizes your cost-per-correct-extraction-with-provenance on your documents, evaluated against your golden set.
Pricing across the tiers spans roughly two orders of magnitude per page. The cheapest credible parser (Mistral OCR 3 Batch at $0.001/page) and the most expensive tier of premium agentic systems (LlamaParse Agentic Plus at $0.056/page; LandingAI ADE Visionary tier) bracket the cost question.

7.1 How to read this chapter

This chapter is not a buyer’s guide that recommends a vendor. It is a map. Buyers’ guides date badly in this category — every quarter at least one vendor releases a major update that changes the ranking — and the goal here is to give you a stable mental model that survives the next twelve months of releases. The vendor names will move. The categories should hold.

Three angles for reading the chapter. If you are a procurement-led reader, focus on the pricing matrix and the per-vendor cost notes. If you are an architecture-led reader, focus on each vendor’s pattern (A, B, or C from Chapter 4) and what they have engineered well around the model layer. If you are a strategy-led reader, focus on the tier structure — the move from “buy a vendor” to “buy a platform” is the strategic shift this chapter is really about.

7.2 The four tiers

The 2026 commercial landscape resolves into four tiers, each with characteristic strengths, characteristic blind spots, and characteristic pricing.

AI-native specialists. Companies founded specifically to do agentic OCR. LlamaParse (part of LlamaIndex), Reducto, Landing AI’s ADE product, Mistral Document AI. These vendors lead the public benchmarks, ship the most aggressive product cadences, and have the most flexible pricing — but their procurement processes are less mature, and their classic-enterprise compliance certifications often lag the hyperscalers and the classic IDP vendors. Most of them are series-A through series-C startups in 2026; some will not survive consolidation.

Hyperscaler APIs. Google Document AI, AWS Textract, Azure Document Intelligence. Mature, broadly compliance-certified, integrated with the rest of the buyer’s cloud stack. Architecturally template-leaning and lagging the agentic frontier on adversarial workloads (especially tables) by a measurable margin. The default pick for buyers who care more about procurement risk than benchmark score, which is a larger fraction of the market than the AI-native specialists like to admit.

Classic IDP, retrofitted. ABBYY, Hyperscience, UiPath Document Understanding, Klippa, Rossum, plus the newer AI-native commercial IDP entrants Nanonets and Docsumo. Strongest HITL infrastructure of any tier; strongest compliance posture (SOC 2, HIPAA, GDPR built into the core product, not retrofitted at sales time); weakest agentic story in the strict five-attribute sense, with notable exceptions. The right tier if HITL is the load-bearing requirement and the agentic loop is secondary.

Platform-integrated agentic OCR. Databricks Document Intelligence is the canonical 2026 example. Not a pure-play OCR vendor; an integrated set of AI functions inside a larger data platform. The strategic significance is bigger than its current market share: a hyperscaler-adjacent platform saying that document reading is the gating constraint on agent quality is air cover for the shift-left thesis (Chapter 6) at the institutional level. Snowflake, Microsoft Fabric, and the major LLM-platform vendors are all building or buying into this tier.

7.3 AI-native specialists

7.3.1 LlamaParse (LlamaIndex)

LlamaParse is the front-runner on the most-cited 2026 benchmark, ParseBench, with their Agentic tier scoring 84.9% overall at approximately $0.012 per page (LlamaIndex and Kaggle 2026). The product offers four tiers — Fast (1 credit/page), Cost-Effective (3), Agentic (10), and Agentic Plus (45) — at a conversion rate of 1,000 credits to $1.25 and 10,000 free credits per month for new accounts.

Architecturally LlamaParse runs Pattern B (agentic loop with reflection) at the Agentic and Agentic Plus tiers, and something closer to Pattern A with light validation at the Fast and Cost-Effective tiers. Strengths: messy enterprise documents, financial filings, scientific papers, and integration with the rest of the LlamaIndex ecosystem (which is itself winning meaningful market share in the agent-orchestration layer). Weaknesses: the system is at its best on text-heavy documents and lags Pattern C specialists on adversarial tables; pricing at the top tier is among the highest in the AI-native category.

The honest framing: LlamaParse is the right pick when you want a single vendor for parsing-plus-orchestration, when your downstream system is built on or compatible with LlamaIndex, and when your documents are text-heavy enough that the Agentic tier (not Agentic Plus) is sufficient. Caveat the ParseBench number with the usual vendor-published warning — the benchmark is LlamaIndex-maintained, and the rule selection inevitably shapes the leaderboard.

7.3.2 Reducto

Reducto is the Pattern C standard-bearer in the commercial landscape. Their hybrid CV-plus-VLM architecture is the most explicit implementation of “specialization beats generality” in the commercial category, and their result on RD-TableBench (~0.90 cell-alignment similarity, ~20 percentage points ahead of the hyperscaler APIs) is the cleanest data point (Reducto AI 2025).

Strengths: complex tables, scanned PDFs with mixed content, handwritten regions, and any workload where layout matters more than text fluency. Their published work on multi-pass agentic OCR (Reducto AI 2026) is the most technically detailed public account from any commercial vendor in 2026 — worth reading regardless of whether you intend to buy.

Pricing runs $0.005–$0.020 per page. SOC 2 and HIPAA available. The honest caveat: RD-TableBench is Reducto-published. Their lead on it is real, but you should cross-validate on a non-Reducto benchmark before signing — OmniDocBench or your own golden set are both useful.

7.3.3 Landing AI — Agentic Document Extraction (ADE)

ADE moved to credit-based monthly subscriptions in 2025–2026, ranging from $39/month for 120,000 credits up to $2,199/month for 900,000 pages (Landing AI 2026). The credit formula — ceil((input_chars/5,000) + (output_chars/1,000), 0.1) — makes the actual per-page cost dependent on document size, which is honest if confusing.

ADE’s 2026 feature set is the most ontology-rich among the AI-native specialists. The product can identify attestations, ID cards, logos, barcodes, and QR codes as first-class entities — not just as “image regions” that downstream code has to interpret. Agentic table captioning, RBAC, HIPAA on Team and Enterprise plans, and one-click Zero Data Retention put ADE in a strong position for regulated workloads.

ADE is the right pick when ontology richness matters (insurance underwriting packets, healthcare records, KYC document review), when HIPAA or ZDR are non-negotiable, and when the documents have a complex enough structure that the broad entity classification pays off. It is not the cheapest tier, and the credit formula needs a few sample runs against your real documents to predict monthly cost confidently.

7.3.4 Mistral Document AI (OCR 3)

Mistral OCR 3 sits at the price floor of credible agentic-grade parsing: $2 per 1,000 pages standard, $1 per 1,000 pages via Batch API, with throughput of up to 2,000 pages/minute per node (Mistral AI 2026). The model identifier mistral-ocr-2512 is callable via API, and self-hosted deployment is available for organizations with strict data-sovereignty requirements.

The internal reported metric is a 74% win rate over Mistral OCR 2 across a customer-document workflow set, particularly strong on forms, handwritten content, and table-heavy documents. Output is Markdown, with tables reconstructed using HTML tags including rowspan and colspan — usable for downstream chunking and embedding without further parsing.

OCR 3 is the right pick when cost-per-page is the binding constraint, when throughput requirements push the hyperscalers’ price-tier into the same range as a specialist, and when self-hosting is required for compliance reasons that exclude the SaaS-only specialists. Architecturally it is closer to Pattern B than to Pattern C — the agentic loop is real but the specialized-tool routing is less elaborate than what Reducto offers.

7.4 Hyperscaler APIs

The three hyperscaler ecosystems — Google Document AI, AWS Textract (plus Bedrock Data Automation), Azure Document Intelligence (plus Azure AI Vision Read) — are mature, integrated, and significantly more nuanced than a single “hyperscaler OCR API” framing suggests. Each is really a family of sub-products at different price points, with different output shapes, and with different positions on the spectrum from “cheap text OCR” to “agentic-flavored extraction.” Buyers who pick the wrong sub-product within the right hyperscaler are a common 2026 failure mode.

This section walks each ecosystem by sub-product rather than as a monolith, because the procurement-relevant choices live at the sub-product level.

7.4.1 Azure — Document Intelligence and AI Vision Read

Microsoft ships document AI through two distinct services that are commonly confused:

Azure AI Vision Read is a standalone OCR endpoint inside the broader Azure AI Vision service. It returns text with reading order; it does not return structured layout or fields. Cost is roughly $1–$1.50 per 1,000 pages. The right pick when you only need clean text — accessibility transcription, search indexing, bulk archive digitization — and do not want to pay for structural extraction you will not use.

Azure AI Document Intelligence (formerly Form Recognizer) is the full document AI product, and it splits into four sub-models worth understanding individually:

Read — $1.50/1k pages. OCR plus reading order plus basic key-value pair extraction. Cheapest tier of Document Intelligence; the right pick when Vision Read is not quite enough but you do not need tables or structural metadata.
Layout — $10/1k pages. Extracts text, tables, paragraphs, section headings, selection marks (checkboxes), and figure regions with bounding boxes. This is the sub-model most relevant to agentic OCR pipelines. It serves the same architectural role as PP-StructureV3 or Granite-Docling in a Pattern C hybrid — a strong layout-detector primitive that downstream stages can route off. Many in-house agentic OCR systems built on Azure use Layout as the parsing layer and add their own reflection loop on top.
Prebuilt models — $10/1k pages. Pretrained extractors for common document types: invoices, receipts, ID documents, business cards, W-2, 1098, contracts, marriage certificates, mortgage documents, health insurance cards. The right pick when your documents are one of the supported types and the extraction schema matches your downstream consumer’s needs.
Custom — $30/1k pages. Train your own extractor on labeled samples. Template models (free training) for narrow stable layouts; neural models (free for the first 10 hours of training, $3/hr thereafter) for variable layouts. The custom-model story is the strongest of any hyperscaler — Azure has invested deeply here and it shows.

High-volume commitments drop pricing meaningfully: at 8 million pages per month, the effective rate falls to around $0.53/1k pages — roughly 50% off list. Free tier covers the first 500 pages per month.

Honest framing: Azure DI in 2026 is the most complete hyperscaler offering for buyers who want to compose their own agentic pipeline from primitives. Layout + Prebuilt + Custom + Azure AI Foundry orchestration gives you most of a Pattern C architecture, with the compliance posture (HIPAA BAA, SOC 2 Type II, FedRAMP High) that classic IDP vendors fought to match.

7.4.2 Google — Document AI and the Layout Parser

Google Document AI is similarly a family of processors, each with its own price point:

Document OCR — approximately $0.65/1k pages for basic OCR. Cheapest tier; text + reading order, no structure.
Layout Parser — $10/1k pages, and the most strategically interesting Google offering in 2026. It uses Gemini under the hood to do layout extraction plus initial chunking in a single call. The output is structured for direct downstream use in RAG and agentic pipelines, which makes it the closest hyperscaler equivalent to the AI-native specialists’ agentic mode. Worth shortlisting for buyers already committed to GCP who want to skip the “compose your own pipeline” engineering.
Form Parser — $30/1k pages below 1M/mo, $20/1k above. Forms plus tables in one call. Older product, still useful.
Specialized processors — pretrained extractors for invoices, US driver licenses, US passports, IRS 1040, W-2, and a dozen other specific document types. Pricing varies; broadly comparable to Form Parser.
Custom Document Extractor — $30/1k pages below 1M/mo, $20/1k above. Train your own. Hosting a deployed custom processor adds $0.05/hour (~$438/year for always-on).
Document AI Workbench — the orchestration layer that lets you chain processors. Increasingly the entry point Google recommends for new agentic workloads.

New customers get $300 in free GCP credits, which goes a long way at these unit costs.

Honest framing: Google’s Layout Parser is the hyperscaler product most under-appreciated by 2026 procurement teams. A team that needs Pattern B parsing and is on GCP can shortlist Layout Parser alongside LlamaParse Agentic and find it competitive on quality at a comparable price point — without the third-party vendor relationship.

7.4.3 AWS — Textract and Bedrock Data Automation

AWS has the messiest sub-product surface, partly because Textract grew incrementally and partly because the newer Amazon Bedrock Data Automation product is now the recommended starting point for new IDP workloads, but Textract remains the right answer for many existing pipelines.

Textract sub-features:

DetectDocumentText — approximately $1.50/1k pages for basic OCR. Text only, no structure.
AnalyzeDocument — feature-stackable, $15–$65/1k pages depending on which features you turn on:
- Forms ($50/1k) — key-value pair extraction.
- Tables ($15/1k) — table structure extraction.
- Queries ($15–$65/1k) — ask questions like “what is the invoice number” and get answers.
- Layout ($10–$25/1k) — paragraph, header, footer, and reading order detection. AWS’s equivalent to Azure DI Layout, added later and less mature. Stack them: a typical real pipeline uses Forms + Tables + Layout together.
AnalyzeID — $2.50/1k. Prebuilt extractor for driver licenses and passports.
AnalyzeExpense — $10/1k. Prebuilt for receipts and invoices.
StartLendingAnalysis — multi-document workflow for mortgage and lending packets. Premium pricing.
Amazon Augmented AI (A2I) integration — the HITL queue layer, well-engineered and the strongest HITL story among hyperscalers.

Amazon Bedrock Data Automation (BDA) — released late 2024, production-ready by 2026 — is AWS’s agentic-flavored answer to LlamaParse Agentic and Azure DI Layout. Per-page pricing is $0.010/page for parsing plus $0.040/page for Custom Output (with an incremental charge if your blueprint defines more than 30 fields). The pricing is flat per document, which makes cost forecasting straightforward. BDA handles classification + extraction through a single multimodal API — and the API itself is multimodal beyond documents (audio, video, image), reflecting AWS’s strategic positioning of Bedrock as the broader unstructured-data platform.

AWS’s own guidance recommends BDA as the starting point for new IDP projects, and Textract for incremental work on existing pipelines. That is the right framing for buyers as well: if you are greenfield, start with BDA; if you have Textract in production already, the migration cost may not justify switching.

Honest framing: AWS in 2026 is the most fragmented of the three hyperscalers and also the most actively evolving. The cost-effective configurations for an agentic workload involve BDA for parsing + reasoning, with Textract feature-stacks as fallback or supplemental tools. Buyers who pick “AWS” without picking a specific configuration usually end up over-paying on Textract features they did not need.

7.4.4 The general framing across all three

Two patterns hold across hyperscaler ecosystems:

The “Layout” sub-product is the load-bearing piece for agentic workloads. Azure DI Layout, Google Document AI Layout Parser, AWS Textract AnalyzeDocument (Layout feature). These are the primitives that practitioners actually use as the parsing layer in custom agentic pipelines, including Pattern C hybrid architectures. The basic OCR tiers (Vision Read, Document OCR, DetectDocumentText) are too thin; the prebuilt/extraction tiers are too narrow.
The agentic story is the newest layer. Google Layout Parser (Gemini-powered) is the most production-credible hyperscaler-native agentic offering. AWS Bedrock Data Automation is close behind and improving. Azure’s agentic story currently relies on chaining Layout + Custom + Foundry orchestration rather than a single packaged agentic product. Watch this race through 2027 — it is the one closing fastest.

The honest framing of the hyperscaler tier overall: these are the right pick for buyers whose procurement processes require an established cloud-vendor relationship, whose compliance requirements rule out smaller vendors, and whose documents fit within the hyperscalers’ design envelope. Choosing the wrong sub-product within the right hyperscaler is the failure mode this section is calibrated to prevent.

A particular failure mode worth naming: teams that pick a hyperscaler because “the procurement is easier” and then discover six months later that the documents they actually want to process are in the gap between what the hyperscaler does well and what the workload requires. The savings on procurement are real; the costs of the wrong architecture are larger. Run the cross-validation on your golden set before committing — and use the right sub-product when you do.

7.5 Classic IDP, retrofitted

The classic IDP vendors — ABBYY Vantage, Hyperscience, UiPath Document Understanding, Klippa, Rossum — have been incumbents in this space for fifteen to twenty years. Their strengths and weaknesses are inverted relative to the AI-native specialists.

Strengths: HITL is native, not retrofitted. The reviewer queue, the reviewer UI, the case-management workflow, the audit trail — these have been the core product for two decades. The agentic specialists are now building this functionality and discovering it is harder than it looks. Compliance posture is similarly mature: ABBYY, Hyperscience, and UiPath all carry the full enterprise compliance certification suite as a default rather than a sales-time question.

Weaknesses: the agentic story is largely retrofitted. The model layer at these vendors is being upgraded to include VLMs and reflection loops, but the surrounding architecture was designed for a different world. The “agentic” features in classic IDP vendor pitches in 2026 frequently fail the five-attribute test from Chapter 3 — typically missing tool-using and goal-driven attributes, or having only weak versions of self-correction.

Per-page pricing in this tier runs $0.020–$0.150, with Hyperscience at the high end of the range. The pricing reflects what classic IDP has always been about: HITL infrastructure and compliance, with parsing as one component rather than the whole product.

A separate sub-tier worth calling out is the AI-native commercial IDP entrants: Nanonets and Docsumo. These vendors are younger than the classic IDP incumbents and are explicitly marketing agentic capabilities. Nanonets prices in workflow blocks ($0.30 per complex AI block, leading to under-$2-per-invoice end-to-end pricing), supports 100+ languages, and reports 34% of Fortune 500 as customers (vendor claim; treat accordingly). Docsumo claims 150+ document types out of the box at 95%+ accuracy, with a free 14-day / 1,000-page trial and tiered subscriptions thereafter.

The honest framing of this tier: choose classic IDP or AI-native commercial IDP when HITL is the load-bearing requirement, when the documents are within the classic-IDP design envelope (invoices, claims forms, KYC, contract abstraction), and when your procurement process is more comfortable with established vendor relationships than with the AI-native startup risk. Do not choose this tier expecting the agentic capabilities to match the specialist vendors at the same price point.

7.6 The platform tier — Databricks, Snowflake, Microsoft Fabric

The most important structural shift in 2026 commercial agentic OCR is the rise of a platform tier distinct from pure-play vendors. Document parsing has been absorbed as a primitive into the major enterprise data platforms — Databricks Document Intelligence, Snowflake Cortex AI, and Microsoft Fabric AI Functions. For buyers already committed to one of these platforms, the procurement story has changed: agentic OCR is no longer a separate vendor relationship but a SQL function or AI Function call billed against existing platform capacity.

Three vendors, three slightly different shapes:

7.6.1 Databricks Document Intelligence

Databricks launched Document Intelligence in 2026 as a set of chainable SQL/AI functions: ai_parse_document (generally available), ai_classify, and ai_extract (Databricks 2026). The integration with Lakeflow, Unity Catalog, and Agent Bricks is tight enough that the product is best understood not as an OCR API but as a primitive in the Databricks data platform.

The publicly reported numbers are aggressive. 5–7× lower cost than comparable VLM-only pipelines for equivalent accuracy. A customer case study (Loopback Analytics, clinical-note extraction) reports approximately 90% cost reduction for the same downstream entity-extraction outcome. The 16% across-the-board performance gain from inserting ai_parse_document upstream of agent frameworks (Chapter 6) is the cleanest single piece of evidence for the shift-left thesis in any 2026 publication.

The strategic significance is larger than the product. A platform vendor at Databricks’ scale saying — with public benchmarks and customer case studies — that document reading is the gating constraint on agent quality is a meaningful institutional endorsement of the shift-left thesis.

7.6.2 Snowflake Cortex — AI_PARSE_DOCUMENT

Snowflake matches Databricks’ move with AI_PARSE_DOCUMENT, a Cortex AI function that extracts text, data, layout elements, and images from documents stored on internal or external stages (Snowflake 2026). The naming convention is not an accident — the two products converged on the same SQL-function shape, the same chainable-with-other-AI-functions architecture, and the same “agentic OCR as a primitive inside the data platform” positioning.

Supported formats: PDF, DOCX, PPTX, JPEG, JPG, PNG, TIFF, TIF, HTML, TXT. Image extraction (added January 2026, in preview) lets AI_PARSE_DOCUMENT emit images embedded in documents — extracted images can be written to stages or passed directly to other Cortex AI functions for further analysis, which is the chaining pattern Snowflake recommends for end-to-end agentic pipelines.

Billing is per page (with HTML/TXT chunked at 3,000 characters per billable page), charged against Snowflake credits. Pricing varies meaningfully by warehouse size and credit rate; the field-reported cost is roughly competitive with Databricks at typical configurations, with the caveat that runaway costs are a known failure mode — there are public 2026 incident writeups of single-query costs in the thousands of dollars when AI_PARSE_DOCUMENT is run against a large corpus without throttling. The cost-monitoring patterns in Chapter 11 apply doubly here.

7.6.3 Microsoft Fabric — AI Functions + Foundry Tools

Microsoft Fabric takes a slightly different shape. Rather than ship a single parse_document function, Fabric integrates Azure AI Document Intelligence (via Foundry Tools) as a callable AI Function inside the Fabric environment (Microsoft 2026). AI Functions in Fabric support multimodal input — images, PDFs, text — and use Fabric authentication, with all usage billed against the customer’s Fabric capacity rather than separately metered.

The practical effect: a Fabric customer running document workflows is using Azure DI Layout (or DI Prebuilt/Custom) under the hood, but through the Fabric interface rather than as a separate Azure subscription. SharePoint Premium (formerly Syntex) is the Microsoft 365 side of the same integration, oriented at end-user document workflows rather than analytics pipelines.

The architectural distinction from Databricks and Snowflake is real: Databricks and Snowflake built their own document AI; Microsoft integrated Azure’s. Both approaches work; both have customers; the choice is downstream of which platform you already use.

7.6.4 The strategic significance

A platform-tier offering is what every cloud-native data team encounters first now when they search “how do I parse documents in [platform].” This changes vendor selection at the procurement level: for buyers already on one of these three platforms, the platform-tier product is the default candidate, and pure-play specialists have to actively justify being added on top.

The honest caveats are familiar across all three: vendor-published numbers, benchmarks maintained by the platform vendor, and architecture that is platform-locked (each product is meaningfully useful only if you are already on the platform). For non-Databricks / non-Snowflake / non-Fabric customers, the platform tier is a data point about where the institutional landscape is heading, not an option you can buy directly. The pure-play specialists are not wrong to be nervous — but they are also not displaced from workloads outside these three platforms, which is most of the addressable market.

7.7 Pricing matrix

The data in data/commercial-players.csv and the per-vendor cost notes above resolve into the following compact summary:

Tier	Vendor / Sub-product	Per-page low	Per-page high
AI-native	LlamaParse	$0.00125	$0.05625
AI-native	Reducto	$0.005	$0.020
AI-native	Landing AI ADE (monthly)	$39/mo	$2,199/mo
AI-native	Mistral Document AI	$0.001	$0.002
AI-native commercial IDP	Nanonets	$0.05/block	$0.30/block
AI-native commercial IDP	Docsumo	(subscription)	(subscription)
Hyperscaler — Azure	AI Vision Read	$0.001	$0.0015
Hyperscaler — Azure	DI Read	$0.0015	$0.0015
Hyperscaler — Azure	DI Layout	$0.010	$0.010
Hyperscaler — Azure	DI Prebuilt	$0.010	$0.010
Hyperscaler — Azure	DI Custom	$0.030	$0.030
Hyperscaler — Google	Document AI OCR	$0.00065	$0.0015
Hyperscaler — Google	DocAI Layout Parser	$0.010	$0.010
Hyperscaler — Google	DocAI Form Parser	$0.020	$0.030
Hyperscaler — Google	DocAI Custom Extractor	$0.020	$0.030
Hyperscaler — AWS	Textract DetectDocumentText	$0.0015	$0.0015
Hyperscaler — AWS	Textract AnalyzeDocument	$0.015	$0.065
Hyperscaler — AWS	Textract AnalyzeID / AnalyzeExpense	$0.0025	$0.010
Hyperscaler — AWS	Bedrock Data Automation	$0.010	$0.040
Classic IDP	ABBYY Vantage	$0.020	$0.100
Classic IDP	Hyperscience	$0.030	$0.150
Classic IDP	UiPath DU	$0.020	$0.080
Classic IDP	Klippa	$0.010	$0.040
Classic IDP	Rossum	$0.020	$0.080
Platform	Databricks Document Intelligence	(platform-tier)	(platform-tier)
Platform	Snowflake Cortex AI_PARSE_DOCUMENT	(platform-tier)	(platform-tier)
Platform	Microsoft Fabric + Foundry Tools	(Fabric capacity)	(Fabric capacity)

Two orders of magnitude separate the floor and the ceiling. The salient point is not which vendor is cheapest; the salient point is that pricing differences across vendors look big until you remember that error rate is the multiplier. A vendor that is four times cheaper per page and has four times the field-level error rate is, on cost-per-correct-extraction, identical to the more expensive vendor. The pricing matrix is only meaningful when paired with the error-rate column, and the error-rate column only becomes meaningful when you measure it on your own documents.

7.8 How to read this landscape

The chapter’s organizing principle, restated as actionable advice:

Do not pick by benchmark rank. Benchmark leaders frequently produce worse production results than benchmark followers on specific workloads. The benchmark is a filter, not a ranking.
Shortlist three vendors across three tiers — one AI-native, one hyperscaler or classic IDP, one Pattern C specialist if your workload is high-stakes. Run all three on your golden set. Decide on data, not demos.
The right vendor is the one that minimizes cost-per-correct-extraction-with-provenance on your documents. This is a measured number, not a marketing number. The only honest way to get it is to run the parsing on a representative sample of your real workload.
Procurement maturity matters more than buyers’ guides admit. A vendor with the right benchmark score but with a six-month security-review timeline can lose to a vendor with a lower benchmark score but a two-week procurement path, simply because the higher-benchmark vendor’s project does not ship within the fiscal-year window. Factor this in honestly.

The chapters that follow build on the same picture. Chapter 8 covers the open-source side of the landscape, which has caught up on parsing quality more than the commercial vendors will acknowledge. Chapter 9 turns the landscape into a decision framework, including the explicit “should I not build this at all” rubric from Section 9.5. Chapter 11 gives you the eval methodology you need to make the vendor-selection numbers honest.

Named takes

The hyperscaler APIs are still where most enterprise procurement lands, and they are still the wrong choice for 2026 agentic workloads. They were designed for templates.

Pricing differences across vendors look big until you remember that error rate is the multiplier. A 4× cheaper parser with a 4× higher field-error rate is the same cost-per-correct-extraction.

The platform tier (Databricks today; Snowflake and Fabric next) is the most important strategic shift in 2026 commercial agentic OCR. The pure-play specialists are not wrong to be nervous.