Why does a frontier model still miss rows in a boring PDF? Adit Abraham, co-founder and CEO of Reducto, told Latent Space the format was never built for agents to read.
Parsing, which is turning visual layout into structured data a model can act on, fails because PDFs preserve print fidelity, not logic. Abraham said: "The original context for PDF as a file format was for printers to be faithful to the original... it wasn't really designed around this sort of problem of agents are going to read it." Reducto processes a little over a billion pages every single month, Abraham said, with customers including Harvey in legal and Basis on tax automation. He added: "We liken it a lot to self-driving car research."
The fix, he argued, is R1.
R1, which is Reducto's in-house model for end-to-end document processing, handles layout capture and last-mile corrections, Abraham said. Structured extraction that historically cost over 4.5 cents per page all in now runs at half that rate with better citations and schema adherence, Abraham said. Reducto later put a public price on that push with r-1 mini at $1 per 1,000 pages for structured output, the company said.
That pricing hits teams that run DIY pipelines on frontier models. Abraham argued those models can win benchmarks with uncapped reasoning time yet silently drop rows and lose recall in production. Datasite's switch from DIY parsing to Reducto shows the same pressure to buy accuracy per dollar. One legal tech customer saw users interact with documents twice as much after switching to Reducto parses, with better answers and citation precision, Abraham said.
Benchmarks remain marketing material, Abraham warned, which is why Reducto used Micro One for an independent test. Watch recall on long tables and multi-hundred-page filings, not just headline accuracy, and whether R1 holds that halved cost at scale. Reducto started the year with one account executive and now has 30 people in sales, Abraham said, with around 40 roles open and capacity as the constraint.
