LandingAI has landed Agentic Document Extraction (ADE) Gen2The new DPT-3 family of models is the basis for a rebuilding its document intelligence stack. Gen1 viewed a document like a list of flat chunks. Gen2 views it as a “tree”, prices the output by characters rather than pages, and grounds each answer to a particular line or word. The LandingAI release is framed around three main themes: affordability; agent-ready outputs and atomic grounds.
Is it deployable? Yes. ADE Gen2 can be downloaded now. Start free for developers in the ADE playground. The Enterprise can use it on the cloud in US and EU, within their VPCs on AWS Azure Google Cloud or Google Cloud or inside Snowflake.
Instead of just one parsing model, there are two.
Gen2 divides parsing in two models to allow the workload to determine the price. DPT-3 Verity It deterministically transcribes documents created digitally and provides a bounding-box and confidence score for each word. This software is designed to handle large amounts of text and form fields. DPT-3 Pro is able to read the page before it even begins, identifies blocks types through signatures and marginalia from figures and tables, then returns them as they were intended. DPT-3 Pro also works with scanned pages and LaTeX math, along with non-Latin characters and handwriting. LandingAI states that DPT-3 Verity is charged at roughly 40% the price of DPT-3 Pro. Also, automated routing between both products will be available in fall 2026.
It’s not the price change that is important, but rather it’s how much you pay.
Each page was 3 credits under DPT-2. DPT-3 allows you to use the DPT-2 system. credit consumption The sum of the page component plus an output character component. DPT-3 Pro charges 1 credit for each page and 0.5 credits for every 1,000 output characters. DPT-3 Verity charges 0.3 credits for each page and 0.2 credits for every 1,000 characters. Standard Tier halve both rates. For example, a 12-page Pro parse that returns 48,120 output characters would cost 36.1 credits in priority tier and half as much on standard. The totals are rounded up to 0.1 credits, and response metadata includes all inputs to the calculation.
Second lever are service levels. Priority can be used when an agent or person is in a waiting state. Standard operates asynchronously and at 0.5x cost, which is suitable for pipelines that can tolerate up to several hours. Playgrounds and synchronous calls run on priority. LandingAI estimates cost savings of 25% to 80% on mixed workloads, and says Verity standard parsing costs less than one cent. These are vendor figures, so you should benchmark your document mix before relying on them.
No chunks, only blocks
You can also find out more about the following: Parse v2 response Three top-level fields are available: Markdown In reading order You can find out more about it here:” You can find out more about the structure by clicking here.. Structure is made up of a node document whose child pages are, and whose child blocks are. The block types are text, table_cells and figures. They also include marginalia (marginals), attestation (attestations), logo, card, scan_code, etc. The semantic ID of each block is in the following form type-indexThe page number is included in the object as the “grounding” element, along with a string of characters that can be inserted into markdown and the normalized bounding area.
The output of Markdown is also standardised. Figures use
[STAMPED][SIGNED]The. [ILLEGIBLE_SIGNATURE] You can also find out more about the following: [ILLEGIBLE_TEXT] As fixed literals. Tables are delivered as HTML default for preserving merged cell.
The power of atomic grounding
Atomic grounding, however, is the most important release feature. Each leaf block contains an atomic_grounding DPT-3 Verity: One entry for each visual line. DPT-3 Pro: one entry for every word. Verity gives a confidence score from 0-1 for each word. The value is computed by the lowest possible character scores in the word. Teams can use this to route transcriptions that are uncertain to review. Pro has left the atomic level grounding of cells empty. Tables now have their own bounding box. The grounding allows Extract V2 citations to be drawn from it, so a field extracted can trace back to an exact word in a particular page. This allows for PII to be redacted by coordinates, and document diffing. It also makes it possible to build UIs that allow reviewers or other users of the system.
