Introducing Light Parse: accurate parsing at a fraction of the cost

TL;DR: Light Parse delivers high-accuracy document parsing for straightforward documents at 50%+ lower cost than other providers, placing it on the cost-accuracy Pareto frontier on both RealDoc-Bench and Databricks OfficeQA Pro.

Light Parse highlights:

• Starts at $0.00625 per page and scales down further with volume discounts
• On RealDoc-Bench, Light Parse placed first among non-agentic parsers, with 90.5% per-field accuracy for document Q&A
• On Databricks OfficeQA Pro, Light Parse placed first among non-agentic parsers
• Best for the majority of high-volume document workflows

Today we’re introducing Light Parse, our latest engine that sets a new standard in accurate, cost-effective document parsing. It provides AI teams with accurate parsing on the majority of documents at 50%+ lower costs than alternatives.

We built Light Parse to equip developers with complete control over every point of the Pareto curve of cost, accuracy, and latency. For the majority of PDFs that don't have handwriting, lengthy tables, or messy checkboxes, Light Parse provides high levels of performance at a fraction of the cost.

Already, Light Parse has helped our customers in industries with high-volume document workflows (e.g., legal AI) reduce their OCR spend in half after migrating from other solutions.

Light Parse performance on open-source benchmarks

We evaluated Light Parse against two open-source benchmarks:

1. Databricks OfficeQA Pro, a third-party agent benchmark created by Databricks
2. RealDoc-Bench, our own open-source agent benchmark comprised of real-world documents

Both benchmarks are openly available so anyone can independently reproduce the results.

Databricks OfficeQA Pro: Light Parse sits on the Pareto frontier of cost vs accuracy

Databricks OfficeQA Pro measures how accurately an AI system can answer exact questions from scanned U.S. government financial documents, including dense historical tables. The test uses the same questions, reader model, and scoring method for every parser, and supplies the relevant pages directly so the results focus on parsing quality. We ran the evaluation against this benchmark, but we did not design it.

On Databricks OfficeQA Pro, Light Parse placed first among non-agentic parsers, reaching 54.9% accuracy at $0.00625 per page, putting it on the cost-accuracy frontier.

If you’re interested in running the evaluation yourself, here’s the Databricks OfficeQA Pro benchmark repository.

RealDoc-Bench: Light Parse sits on the Pareto frontier of cost vs accuracy

RealDoc-Bench QA measures whether an agent can answer questions correctly from parser output, with separate per-field and per-question scores. The benchmark corpus includes real-world mortgage, finance, supply chain, and healthcare documents (Hugging Face).

On RealDoc-Bench, Light Parse achieved 90.5% per-field accuracy and 82.7% per-question accuracy. This makes it the most accurate parsing engine with no agentic configurations enabled.

It spiked in particular on mortgage documents, reaching 94.4% per-field accuracy and 88.1% per-question accuracy.

You can review RealDoc-Bench’s methodology and datasets and run the benchmark yourself (GitHub repo).

How we built Light Parse

Light Parse runs smaller, faster versions of our table and form models, and skips some model calls entirely that aren't necessary unless maximum performance is required.

We’re able to lower parsing cost while preserving performance because of our upfront investment in building a new state-of-the-art layout model from the ground up. With a strong layout model powering our parsing engines, we’re able to deliver more deterministic controls over document processing.

When should you use Extend Light Parse vs. Performance Parse?

Light Parse is well suited for text-heavy files when document volume and cost are important. These files can include reports, agreements, and digital documents in Germanic and Latin-based languages.

Performance Parse is better suited for large tables, dense forms, checkboxes, messy handwriting, non-Latin based languages, and low-quality scans.

*Additional volume-based discounts available.

At pay-as-you-go rates, one million pages cost $6,250 with Light Parse or $25,000 with Performance Parse, excluding any additional agentic configurations.

Both engines can be configured using the same Parse API. You can select the engine with one configuration field:

{
  "file": {
    "url": "https://example.com/document.pdf"
  },
  "config": {
    "engine": "parse_light",
    "engineVersion": "1.0.0"
  }
}

What’s next?

Soon, we're releasing our Auto mode, which intelligently routes each page to Light Parse or Performance Parse based on a classifier model we've tuned to detect predicted page complexity.

This will help teams process documents efficiently and effectively within their latency and cost constraints while preserving the output quality their agents need.

Get started

Light Parse is generally available. Give it a try on your documents today for free: https://dashboard.extend.ai/.