# Reducto launches r-1 at 1 cent per page for document parsing > Source: > Published: 2026-09-01 16:38:00+00:00 # Reducto launches r-1 at 1 cent per page for document parsing **Adit Abraham and Raunak Chowdhuri launched Reducto's preview document model after the company raised $108M, pricing r-1 at 1 cent per page.** By [RuntimeWire Staff](/author/runtimewire-staff) ยท Published Primary source: [PR Newswire](https://www.prnewswire.com/news-releases/reducto-unveils-a-frontier-parsing-model-that-makes-the-worlds-hardest-documents-ai-ready-for-1-a-page-302866077.html) ## Why it matters Document ingestion is becoming a model layer of its own. If Reducto's claimed gains hold under independent testing, AI builders could replace multi-tool parser stacks with a cheaper API. [Reducto](https://reducto.ai/?ref=runtimewire), the document-intelligence startup founded by [Adit Abraham](https://x.com/aditabrm?ref=runtimewire) and [Raunak Chowdhuri](https://x.com/raunakdoesdev?ref=runtimewire), launched r-1 in preview on September 1 at a company-advertised price of 1 cent per page. It is positioning the model as a replacement for the OCR, layout and vision-language components often assembled into document pipelines. Abraham and Chowdhuri built Reducto around the premise that parsing errors undermine the AI systems using the resulting data. In a [September 1 announcement](https://www.prnewswire.com/news-releases/reducto-unveils-a-frontier-parsing-model-that-makes-the-worlds-hardest-documents-ai-ready-for-1-a-page-302866077.html?ref=runtimewire), Reducto said its early preview reduced parsing errors by up to 20% versus its legacy agentic pipeline. Reducto says r-1 costs about 1 cent per page and can reduce all-in processing costs by up to six times versus the multi-tool pipelines it is designed to replace. These figures come from Reducto's evaluations, and the launch materials do not provide an independent assessment of the preview model. The price is the sharper part of the launch. Reducto's [r-1 announcement](https://reducto.ai/blog/parse-r-1-model?ref=runtimewire) says r-1 costs 1 cent per page. The company says r-1 handles layout, reading order, tables, figures, handwriting, formatting and citations through a full-page model. That gives Abraham and Chowdhuri a straightforward sales pitch: stop maintaining a parser stack and send the document workload to Reducto. ### A memory product led to the ingestion problem Reducto's first product was a long-term memory system for language models. The founders later concluded that document ingestion, rather than retrieval or inference, was a bottleneck in many enterprise AI systems and pivoted toward document parsing. Reducto needed to read uploaded files before its memory product could use them, and the available parsers kept losing tables, mixing columns or breaking the reading order. A simple document-segmentation tool built as a weekend project drew stronger interest than the memory product, according to the [First Round account](https://review.firstround.com/reductos-path-to-product-market-fit/?ref=runtimewire). Abraham had studied computer science at MIT, conducted machine-learning research at the MIT Media Lab and worked on ads and search products at Google. Chowdhuri had built a computational-chemistry business. Their initial division of labor carried into Reducto. According to the [First Round interview](https://review.firstround.com/reductos-path-to-product-market-fit/?ref=runtimewire), Abraham leaned toward product, customers and sales, while Chowdhuri led the deeper technical work. R-1 addresses the ingestion failures the founders encountered. A capable language model can still return the wrong answer when a financial table loses a column, a contract drops a strikethrough or a scanned form assigns handwriting to the wrong field. ### One model takes over the pipeline Reducto describes its broader platform as a hybrid architecture combining layout-first computer vision, vision-language-model review and Agentic OCR multi-pass correction. Its documentation says Parse handles OCR, parsing and [post-processing](/models/fal/Post-Process) such as chunking. Separate [r-1 documentation](https://docs.reducto.ai/parse/r-1?ref=runtimewire) describes the preview model as a unified, full-page option within that platform. The model reads scans and handwriting, resolves columns and sidebars, interprets merged table cells and nested headers, and retains formatting such as lists, underlines and strikethroughs. It returns page-relative bounding boxes so developers can trace an output back to its location in the source document. R-1 remains in preview. Existing Reducto customers can access it through a configuration flag in the company's Parse API. Developers can also augment r-1 with custom prompts through agentic processing. ### Testing still falls to customers Reducto's published claim of up to 20% fewer parsing errors lacks enough disclosed methodology for an independent reproduction. The launch materials do not provide sufficient detail about sampling, annotation, error severity or baseline configurations. Customers will also bring narrower document sets where tables, handwriting, charts or latency carry different weights. ### $108M is funding the model bet Reducto has raised $108 million since its 2023 founding. Its October 2025 [$75 million Series B](https://reducto.ai/blog/reducto-series-b-funding?ref=runtimewire) was led by [Andreessen Horowitz](https://a16z.com/?ref=runtimewire), with Benchmark, [First Round Capital](https://www.firstround.com/?ref=runtimewire), BoxGroup and [Y Combinator](https://www.ycombinator.com/?ref=runtimewire) participating. Reducto says its platform has processed more than 5 billion pages and identifies teams at Harvey, Scale AI and Vanta as users on its [About page](https://reducto.ai/about?ref=runtimewire). The volume and customer names are Reducto's disclosures. Revenue, margins and r-1 adoption were not established in the launch materials. Abraham and Chowdhuri have spent three years turning a weekend document-segmentation project into Reducto's core business. R-1 puts model research and lower per-page pricing at the center of that wager: document ingestion deserves its own trained models, and the winning parser will be judged on the errors it prevents and the infrastructure it lets developers delete.