Unstructured, whose open source software recently surpassed 100 million downloads, has launched new models for Transform, its document parsing solution for turning PDFs, scans and other complex documents into data AI agents can act on. In testing on 224 real-world business pages, GPT-6, Claude and Gemini produced up to 2.3 times as many table errors as Transform and invented about 50% more content that was never in the document.
Unstructured has launched new models for Transform, its document parsing solution for AI agents.
In testing on 224 real-world business pages, GPT-6, Claude and Gemini produced up to 2.3 times as many table errors as Transform.
The frontier models also invented about 50% more content that was never in the document.
Transform costs $0.015 a page, while GPT-6 costs about 7.5 times as much on the same run.
On RD-TableBench, Transform scored 0.899 using Reducto's own grading script, and on forms it finds 89% of fields.
The new models are live in Transform now with 10,000 free pages at transform.unstructured.io.
"RAG needed mediocre data. Agents need extraordinary data," said Brian Raymond, founder and CEO of Unstructured. "At 99% accuracy per step, a 50-step agent finishes clean 60% of the time. At 99.9%, it finishes clean 95% of the time. That extra point nine lives in the tables, forms and scans other tools get wrong. We built Transform to get them right. Every parser wins its own benchmark, ours included, so we'll prove it on your documents."
Tables, forms and scans are where AI fails quietly. A decimal point drops out of a table cell, and $1,407.77 becomes $140,777. An empty checkbox on an insurance claim is read as checked. A blank field on a loan application comes back filled with a hallucinated guess. When a person reviews the output, someone catches the mistake. When an agent acts on it, the error flows straight into a payment, a decision or an answer to a customer.
The 224 pages come from SCORE-Bench, Unstructured's open benchmark of real-world business documents, where Transform ranked first on every table-accuracy measure.
One call to Transform returns tables as real grids with merged headers intact, forms as label-and-value pairs with handwriting and checkbox states preserved, and scanned PDFs with no text layer as clean text, no separate OCR step needed. Every element comes back labeled by type, with its page number and coordinates, so an agent can cite exactly where an answer came from.
In Transform's workbench, users can get a recommended extraction schema from a sample document, simplifying schema iteration. Developers can choose between Balanced, the default parsing strategy, or Best for scans, handwriting and dense tables.
About Unstructured
Unstructured turns the files enterprises run on, from contracts and claims to filings, forms and scans, into data AI agents and applications can use. Its software has been downloaded more than 100 million times and is used by 87% of the Fortune 1000, including Amazon, Bank of America, Citi, Google, Humana, JPMorgan Chase and McKinsey. Its products include Transform, for parsing and extracting documents through an API; Pipelines, for keeping document data current across systems; and Foundation, a single MCP server for agentic search over enterprise data. Founded in 2022, Unstructured is headquartered in Sacramento, California.
