LandingAI has officially released Agentic Document Extraction (ADE) Gen2, a complete overhaul of its document intelligence stack powered by the new DPT-3 model family. The most immediate impact for enterprises is financial. Under the previous DPT-2 system, users paid a flat rate of 3 credits per page regardless of content density. Gen2 shifts to a hybrid pricing model based on page count plus output characters. For high-volume, simple text tasks using the DPT-3 Verity model on the standard tier, LandingAI claims costs can drop below one cent per page. Even complex mixed workloads could see savings between 25% and 80%, though dense pages with heavy character output might actually cost more than before.
The technical architecture has shifted from flat chunks to a hierarchical tree structure. Gen2 splits parsing duties between two specialized models. DPT-3 Verity handles digitally created documents deterministically, providing word-level grounding and confidence scores for every single word. It charges roughly 40% of what the heavier DPT-3 Pro model costs. Pro is designed for difficult inputs like scanned pages, handwriting, non-Latin scripts, and LaTeX math. It analyzes page layout before reading words, detecting block types such as tables, figures, and marginalia. While automated routing between these two models is planned for fall 2026, developers must currently choose the right tool for their specific data type.
Precision is the other major upgrade. Every extracted field now traces back to a specific visual line or word via atomic grounding. This allows for exact coordinate-based PII redaction and reliable document diffing, moving beyond approximate matches. The output format is also stricter, using standardized Markdown with isolated description tags for figures to prevent transcription errors. However, this power comes with a migration tax. Gen1 client code is incompatible with Gen2 endpoints. Teams will need to refactor their integrations to handle the new document-node tree structure and block-based responses. It is a significant break, but one that enables far more robust agent-ready outputs.
