LandingAI ADE Gen2: Per-Character Pricing and Atomic Grounding

LandingAI has shipped Agentic Document Extraction Gen2, a full rebuild of its document parsing stack, generally available now. The release introduces two new models, DPT-3 Pro and DPT-3 Verity, replaces the old flat per-page billing with a combined page-plus-character rate, and adds word-level grounding that traces every extracted field back to a specific location on the page. LandingAI projects cost reductions of 25% to 80% on mixed workloads, though the character component means dense pages can cost more than before.

What happened

Detail	Fact
Availability	Generally available now; free playground for developers
Old pricing	Flat 3 credits per page (DPT-2)
DPT-3 Pro (priority)	1 credit per page + 0.5 credits per 1,000 output characters
DPT-3 Verity (priority)	0.3 credits per page + 0.2 credits per 1,000 output characters
Standard tier discount	0.5x the priority rate; runs asynchronously
Verity cost vs. Pro	Roughly 40% of Pro’s credit cost
Example parse cost	12-page Pro parse returning 48,120 characters: 36.1 credits on priority
Projected savings	25% to 80% on mixed workloads; under one cent per page with Verity on standard (vendor figures)
Automated model routing	Planned for fall 2026
Deployment options	US/EU cloud, VPC on AWS/Azure/Google Cloud, Snowflake, on-premises, air-gapped

Gen1 parsed documents as a flat list of chunks. Gen2 returns a tree: a document node contains pages, each page contains typed blocks. Block types include text, table, table_cell, figure, marginalia, attestation, logo, card, and scan_code. Every block carries a stable semantic ID, a page reference, and a normalized bounding box. The old client code does not work with Gen2 endpoints, so any existing integration needs a migration.

How the two models differ

DPT-3 Verity is built for digitally created documents. It transcribes text deterministically, attaches a confidence score per word (computed as the lowest per-character score in that word), and returns a bounding box for every word. It suits high-volume jobs: structured text, tables, and simple form fields.

DPT-3 Pro reads page layout before individual words. It classifies block types across tables, figures, marginalia, and signatures, returns them in reading order, and handles scanned pages, handwriting, non-Latin scripts, and LaTeX math. Pro grounds to the line level rather than the word level, and it leaves cell-level atomic grounding empty for tables.

Why the pricing change matters more than the new models

The move from flat per-page billing to a page-plus-character model is the most practically significant change. A sparse document with few output characters now costs less than it did under DPT-2. A dense document with many output characters may cost more. LandingAI’s own example: a 12-page Pro parse returning 48,120 characters comes to 36.1 credits on priority, which is lower than the old 36-credit flat rate only by a fraction. On standard tier (0.5x), the same job costs roughly 18 credits.

Synchronous API calls always bill at the priority rate, and the playground runs on priority too, so developers testing in the playground will not see standard-tier pricing until they switch to async pipelines. The response metadata reports every input to the billing calculation, which at least makes the numbers auditable.

LandingAI’s projected 25% to 80% savings should be treated as a vendor estimate. Benchmark your actual document mix before committing. If you are processing dense financial statements or legal contracts, the output character component could push costs above Gen1 levels for those specific jobs.

What atomic grounding actually enables

Atomic grounding is the piece that has the most practical consequence for teams building on top of parsed documents. Every leaf block carries an array of grounding entries: one per visual line with DPT-3 Pro, one per word with DPT-3 Verity. Each entry includes the page number, a range into the markdown string, and a bounding box.

This makes several previously approximate tasks concrete:

• PII redaction by coordinate: mask a field at a known pixel location rather than searching text for a pattern.
• Document diffing: compare two versions of a document at the block level, not just as raw text.
• Reviewer UIs: highlight exactly which word on which page drove an extracted value.
• Confidence-based routing: Verity’s per-word confidence score gives a signal to send uncertain transcriptions to human review.

Extract V2 (LandingAI’s field extraction layer) draws citations directly from grounding, so any extracted field traces to a specific word on a specific page. Table cells also carry their own bounding boxes now, though Pro omits cell-level atomic grounding.

Output format changes worth knowing

Markdown output is standardized. Figures use a custom element style with generated prose inside description tags, keeping model-generated commentary separate from transcribed text. Attestations (stamps, signatures) emit stacked labels such as [STAMPED][SIGNED], with fixed literals [ILLEGIBLE_SIGNATURE] and [ILLEGIBLE_TEXT] for unreadable content. Tables default to HTML output to preserve merged cells.

Why it matters

Document extraction is plumbing for a large share of enterprise AI workflows: invoice processing, contract review, compliance reporting, form digitization. Most current approaches still pass raw PDF text to a language model and hope for the best. Grounded, structured output with bounding boxes is a meaningful step toward pipelines that can actually be audited and corrected.

For teams building AI integrations that touch documents, ADE Gen2’s block tree and per-word grounding reduce the gap between “it mostly works” and “it works and I can prove it.” That matters for any regulated industry or any workflow where a human has to sign off on the output.

Our take

The pricing model is more honest than a flat page rate, but also more complex to forecast. The character component rewards sparse outputs and punishes dense ones. Before migrating, pull a representative sample of your documents, run them through the playground, and read the billing metadata that comes back. Do not rely on the vendor’s 25-80% savings projection for your specific workload.

The atomic grounding feature is genuinely useful and not widely available at this specificity. If you are building anything that needs a human reviewer to verify AI-extracted data, word-level bounding boxes with confidence scores are a real capability gain. The fall 2026 target for automated Verity/Pro routing is also worth watching: it could simplify cost optimization significantly if it lands on schedule.

Gen1 code will not run on Gen2 endpoints. If you are already using ADE, plan the migration carefully. The response schema has changed substantially, and the block-tree structure requires updates to any downstream parsing logic. If you need help scoping that work, our team is available to talk through the integration.

For context on how AI tools are being applied to real business problems, see our coverage of Claude-powered AI in government services as another example of structured AI output needing auditability.

Source: Marktechpost