Introducing Agentic Document Extraction, 2nd Generation

The LandingAI team is thrilled to share with existing customers an overview of all the innovations bundled into Agentic Document Extraction (2nd Generation). This post will be updated every few days across July as we complete fast-follower features and publish more resources to help you understand everything available to you and how to get started using it.

Rethinking Document Processing for the Agentic Era

Agentic Document Extraction is now faster, more accurate, and more efficient than ever before. From a customer point of view, it delivers higher performance while optimizing cost for every document. Customers are getting a platform with more agent-friendly outputs and more detailed citations that is affordable at scale.

Going forward Agentic Document Extraction will be powered by the DPT-3 family of models. This new foundational model is built from the ground up to read documents the way a careful human does, starting with overall layout and structure and ending with individual lines and table cells. Then all that detailed information is codified in a redesigned API response that AI agents love.

Traditional document processing was built for a pre-agent world. LandingAI's Agentic Document Extraction is built for the agentic era. This is not an incremental update. It is a fundamental rethink and rebuild of how AI understands and works with documents in the agentic era.

Three Themes + A Whole Lot More

Over the past year, you told us exactly what you need from document processing to run your most demanding workloads. You asked and we listened, building capabilities to address cost, capabilities for pinpoint traceability at scale, and capabilities to unblock use cases such as redaction.

This release has 3 thematic messages to help you understand the new capabilities as a collection. It also has a long list of point improvements in the Overflow section that don't group neatly into a theme. Look for links to more detailed resources where available. Everything discussed here is also documented in the official change log and migration guide

The themes:

1. Affordability at scale: You can now automate your whole document mix accurately and economically.
2. Agent-ready outputs: Output built for AI agents to consume, not just for people to read.
3. Atomic citations: Line-level and word-level citations open up many new use cases and downstream applications.

1. Affordability at scale

High accuracy in document intelligence is table stakes. After accuracy is achieved, the next two metrics are typically cost and speed. Agentic Document Extraction Gen2 is built to make your entire document mix - the light pages and the heavy ones - economical to automate. We have achieved this via multiple pricing levers.

That claim requires an explanation and proof. First the explanation...

1A. Variable pricing based on content complexity

Complexity-based pricing for Parse is the new core pricing philosophy. When parsing with the DPT-3 models you pay for the number of characters returned, not the number of pages in the document. This means that a content-light page costs the minimum and a dense, content-heavy page costs more. Every Parse response reports the output character count and the final credit cost, so you can see exactly what each page cost and why.

It's worth noting that Extract pricing philosophy is not changing. Extract has always been complexity-based considering both input and output characters. More to say about the impact of all the changes taken together on Extract credit usage. Bookmarking it for now.

1B. Save money by waiting a little longer

Service tiers are new and are the single largest factor in price. The service tier sets how quickly you get each job back. Priority is for when a person or an agent is waiting on the result. Standard is the default. Standard runs asynchronously and suits production pipelines where a response in the range of minutes to hours is acceptable. A lower-cost Batch tier handles high-volume work that can wait hours to days.

Relative pricing is 1.0x for Priority, 0.5x for Standard and 0.375x for Batch (coming soon).

Note that the ADE Visual Playground uses the Priority service tier since a person is waiting on the result. Cut credit usage in half by switching to API-based usage and using the default, Standard tier.

1C. Select the least expensive, capable model

DPT-3 will grow into a family of models starting with DPT-3 Pro and followed by DPT-3 Fast. Users will have the ability to select which model should handle their parsing work based on the complexity of the document. By the Fall of 2026, users will also have the ability to select intelligent routing between models. This is cost-optimization designed in which you don't have to build yourself.

DPT-3 Pro is the high-accuracy model capable of handling extreme document complexity. It is prepared to read your handwriting, deal with forms that have inconsistent checkmarks, work across Chinese and Arabic languages, and parse scanned documents and photographs of documents just as well as digital PDFs. It is the full-capability model.

DPT-3 Fast is a high-accuracy model that offers lower cost for digital text and tables in Latin scripts. Pricing for this model relative to DPT-3 Pro has not yet been set. Customers that process basic documents in English, Spanish, French, German etc. will find that DPT-3 Fast offers them significant cost savings. DPT-3 Fast is designed to compete on price with low-cost OCR models while offering additional features which those models lack.

Recap of what's new

Pricing philosophy

• Complexity-based pricing for Parse. Parse charges variable credits based on the number of characters returned. Users have control over which pages and which block types they want included in the output. [GA now]

Models

• DPT-3 Pro, a high-accuracy model capable of handling extreme document complexity. [GA now]
• DPT-3 Fast, a high-accuracy, lower cost model for digital text and tables. [Preview in August 2026]
• Auto-select, an intelligence capability to route between DPT-3 Pro and DPT-3 Fast automatically. [Preview in Q3/Q4 2026]

Service Tiers

• Parse offers two service tiers with Standard priced at 0.5x compared to Priority. [GA now]
• Extract offers two service tiers with Standard priced at 0.5x compared to Priority. [GA now]
• Batch service tier for Parse and Extract priced at 0.375x [Fast Follower]

Why it matters

The economics often decide which workflows organizations choose to automate at all. Processing pages via agentic document APIs can't cost more than the manual steps being replaced. The new pricing that counts the content on the page and offers tiers aligned with urgency makes it possible for LandingAI to service all your document processing needs from simple to complex.

The payoff on typical mixed workloads is roughly 25 to 40% lower per-page cost with the features available today (July 21), rising to up to roughly 80% lower cost once DPT-3 Fast and intelligent routing between models is in place.

For a deeper dive into pricing, read Pricing Core Concepts and review the worked pricing examples on real documents, which show 40%, 43%, 58%, and 62% cost reductions on four basic use cases.

Go deeper into the details

• Understanding pricing core concepts
• Compare v1 and v2 pricing on real documents and use cases
• Understanding service tiers
• Complete credit consumption documentation

2. Agent-ready outputs

An agent is only as reliable as the context it receives. Agentic Document Extraction Gen2 produces the kind of output agents need. The output is built for machines to consume, meaning that it is predictable and highly structured. The DPT-3 API response is genuinely new and rebuilt from the ground up with AI agents in mind. The JSON output from the API has been completely redesigned to be highly standardized and easy for agents to understand.

For existing customers this means faster delivery of use cases across the enterprise with the help of coding agents. But, it also means setting aside the time to make the switch to the new API response structure.

Your code from ADE Gen1 will not work with Gen2. Hopefully the pricing improvements above have already motivated you to migrate. If you are not convinced yet, this section will introduce you to how we think about agent-ready Parse outputs and why they matter.

2A. Logical, hierarchical structure

With this release we are introducing blocks and retiring chunks. A block is a logical grouping of related content on a page. It could be a table, a paragraph of text, or a bar chart together with its title, legend and caption. Pages contain one or more blocks. The structure is pages → blocks → lines → cells or pages → blocks → lines → words.

Also new is that every block carries a stable ID based on its type and we use Markdown span pointers for localization on the page.

2B. Simplify and standardize markdown outputs

The Markdown itself is now much more highly standardized. This means that customers won't have to plan for multiple edge cases and can expect a conforming response each time.

The most notable changes are for Figures, Attestations, Checkboxes and Tables.

• Figures use standard image syntax with subtype labels ([Chart], [Diagram], [Photograph], [Illustration]).
• Attestations are structured as Signed or Not Signed. The verbose descriptions of the past that added no value are gone.
• Checkboxes are standardized, including checkboxes inside tables.
• Tables are delivered as HTML, which preserves structure that plain Markdown flattens.

Recap of what's new

Response Structure

• New REST Parse API and response structure for DPT-3. Available via the v2 endpoints. See Parse API response [GA now]
• New REST Extract API and response structure. Available via the v2 endpoints. See Extract API response  [GA now]
• Hierarchical layout model (pages, blocks, lines, table cells) with stable Block IDs and Markdown span pointers back to source; Blocks replace Chunks. [GA now]
• More compact output. Dropping verbose descriptions and non-standard tags means the same document generally returns fewer characters. Cleaner for agents to consume and, because Parse pricing tracks output characters, cheaper too. [GA now]

Standardization

• Consistent key–value pairs, intact and predictable. [GA now]
• Standardized Markdown for figures: standard image syntax with subtype labels ([Chart], [Diagram], [Photograph], [Illustration]). [GA now]
• Standardized Markdown for checkboxes: one consistent representation, including checkboxes inside tables. [GA now]
• Standardized Markdown for attestations: structured cleanly as Signed or Not Signed. [GA now]
• Tables are delivered as HTML by default with option to select Markdown instead [GA now]

Generative vs Literal Outputs

• For figures, the literal transcription of the text on the figure is kept separate from the model's generated description of it. Customers always know which is source text and which is AI-generated. [GA now]

Why it matters

Structured, grounded, predictable output is what lets an agent act on a document without a person re-reading it first. When every value has a stable ID and a pointer back to its place on the page, downstream systems can retrieve, reason, and cross-reference without guessing. And they behave consistently across runs. Standardized Markdown puts more signal in fewer tokens, which lowers cost and reduces the chance an agent trips over formatting differences.

Go deeper into the details

• Deep dive on structured attestations and the Signed or Not Signed model, see Detecting signatures, stamps, and seals.

3. Atomic citations, and what they unlock

Trust is what makes scale safe. A system can read a million pages, but if you can't check any single answer against its source, you can't put it near a regulated workflow. Citations tie every extraction back to its exact place in the source. In Gen2, DPT-3 Pro makes them atomic: grounded at the smallest structural units of a document rather than at the level of a whole page or paragraph.

Today, table-cell citations work exactly as they did in Gen1, with every cell mapped to its position. The DPT-3 family adds line-level citations. For a text block this means a bounding box for the whole text block plus a box per line. Coming soon, customers will be able to work at the word level or above and control the citation level they want for a given use case, trading detail against cost. Word-level citations and that configurability are on the way.

These citations live inside Extract V2, the service that pulls the specific fields you define out of a document into structured JSON. Extract V2 keeps the same infinite, fully custom schema and the same output formatting options that ADE already offered, so anything you built against them carries straight over, and it's available through a new Extract Jobs API for submitting extraction jobs and retrieving structured results at scale.

What's new

• Line-level citations for text blocks (a box per Block plus a box per line; covers attestations, cards, and logos). [GA now]
• Table-cell citations, unchanged from Gen1. [GA now]
• New Extract Jobs API for submitting extraction jobs and retrieving structured JSON results at scale. [GA now]
• Extract V2 keeps the same infinite, fully custom schema and output formatting options available before. [GA now]
• Word-level citations. [Fast Follower, targeted end of summer]
• Configurable citation levels: choose the granularity your use case needs (word level and above), trading detail against cost. [Fast Follower]
• The capabilities atomic citations unlock: human verification, PII/PHI redaction, document comparison and version-diffing, in-place editing, and field-level extraction confidence. [Fast Follower]

Why it matters

Because every answer knows exactly where it lives on the page, atomic citations are the foundation for a class of capabilities that couldn't be built on DPT-2. Rapid human verification lets a reviewer jump straight to the source region for any extracted value. PII and PHI redaction can be localized to the exact box. Document comparison and version-diffing work at the field level, in-place document editing becomes possible, and every field can carry an extraction confidence. Grounding at this level of precision is what makes each of these buildable, and letting customers choose their citation level means they pay for exactly the precision a workflow requires. It's why the citation work is the foundation of the release. The pattern to hold onto: autonomous enough to read at scale, auditable enough to trust the results.

Overflow: More in this release

A few other things shipping alongside the three headline changes:

• New SDK version carrying the v2 APIs. [GA now]
• Parse Jobs API for async parsing at scale. Submit large documents for asynchronous parsing — up to 6,000 pages or 1 GiB per PDF — and retrieve results when they're ready. [GA now]
• Unified async job model. Parse Jobs and Extract Jobs now share a single response envelope and job-id format, with the option to store job output for later retrieval. [GA now]
• Standardized error handling. Consistent error responses in a {"code", "message"} shape, with validation failures returning HTTP 422. [GA now]
• Partial-success reporting for Extract. Extract surfaces partial results with warnings and schema-violation details, returning HTTP 206 on synchronous requests that partially succeed — so pipelines can handle incomplete results gracefully. [GA now]
• Correct handling of rotated images. EXIF orientation is now respected, so bounding boxes line up with the document as you see it — important for scans and phone photos. [GA now]
• Redesigned Usage dashboard. Daily and monthly consumption charts plus per-job logging, so you can see exactly where credits go. [GA now]
• Transparent credit pricing. $1 buys 100 credits on Explore and Teams plans, in both the US and EU regions. [GA now]
• Separate documentation. Gen2 is different enough that it ships with its own docs. [GA now]
• New Playground URL at ade.landing.ai. Update your bookmarks. [GA now]
• Enterprise deployment options. In addition to US and EU cloud, choose from your own VPC on AWS, Azure, or Google Cloud, Snowflake, or on-premises, including air-gapped environments. [GA now]
• Improved form handling, especially checkbox detection, including checkboxes inside tables. [GA now]

For the full list of changes, see the changelog.

A new foundation, and a practical place to start

ADE Gen2 is a new foundation for the agentic era: autonomous enough to do the reading at scale, auditable enough to trust with the results. A new model, a new API, pricing that makes the whole document mix economical, and grounding that ties every answer back to the page.

This is bigger than automating paperwork. As AI moves from answering questions to executing work, documents become the foundation, and agents can only make good decisions when the information they act on is accurate, grounded, and trustworthy. That is the agentic document processing infrastructure ADE Gen2 is built to be.

ADE Gen2 is available today. Get started free at ade.landing.ai.