Light Parse sits on the cost–accuracy Pareto frontier in evaluations on RealDoc-Bench and OfficeQA Pro, at $0.00625 per page

Extend, a document processing platform for developers building AI applications and agents, today highlighted results for Light Parse on two open-source benchmarks. The generally available parsing engine achieved 90.5% field-level accuracy on RealDoc-Bench and 54.9% question-answering accuracy in Extend’s evaluation on Databricks OfficeQA Pro. At $0.00625 per page, Light Parse gives developers a cost-efficient way to process documents for AI applications.

Light Parse is well-suited for the majority of document workloads, especially text-heavy, high-volume documents like reports, agreements, and contracts.

Benchmark performance

Both benchmarks test whether an AI system can answer questions correctly using a parser’s output. This measures whether document processing preserves the information and relationships an agent needs to complete real tasks, such as finding a payment amount under the correct table heading or identifying a term in a mortgage document. Scoring those answers against verified ground truth connects parsing quality directly to downstream AI performance.

RealDoc-Bench. On RealDoc-Bench, Light Parse achieved 90.5% per-field accuracy and 82.7% per-question accuracy, which Extend reports as first place among non-agentic parsers and placement on the cost-accuracy Pareto frontier at $0.00625 per page. Extend’s Performance Parse scored 95.7% per-field accuracy at $0.020 per page on the same benchmark view.

Databricks OfficeQA Pro. On Databricks OfficeQA Pro, which measures exact-question answering from scanned U.S. government financial documents (including dense historical tables), Light Parse placed first among non-agentic parsers and places on the cost-accuracy frontier.

Cut parsing costs by over 50% with control over every document
Light Parse gives developers a way to scale document processing while keeping costs predictable. Light Parse builds on Extend’s proprietary layout model to process documents containing text, forms, and tables. It uses smaller, faster table and form models and skips unnecessary model calls, reducing cost while preserving the information AI applications need. Customers in industries including legal AI have already cut their OCR spending in half after switching from other solutions.

Developers can select Light Parse or Performance Parse via API or the Extend Studio , matching the engine to each document’s requirements. Complementing these engines is Extend’s work on Auto mode: routing each page based on predicted complexity, with straightforward pages handled by Light Parse and more demanding pages by Performance Parse.

Light Parse is generally available, try it on your documents today.

About Extend

Extend is an Applied AI lab that provides models, APIs, and infrastructure to enable developers and agents to make every document process programmatic. Its end-to-end platform delivers API primitives and tooling to parse, extract, split, edit, run evals, and orchestrate workflows to turn unstructured documents into structured data that agents can reason over and act on. Companies including Brex, Mercury, Opendoor, Amgen, First American Financial, C.H. Robinson, Flatiron Health, and more use Extend to power document ingestion in production.

View source version on businesswire.com: https://www.businesswire.com/news/home/20260922514533/en/