Reducto Releases r-1 Document Parsing Model With 20% Fewer Errors

Last week, Reducto announced the release of r-1, its inaugural model in a new document parsing family. This model features a rewritten architecture designed to replace the company's previous multi-stage agentic OCR process with a single, full-page pass. Reducto claims that r-1 surpasses its most powerful legacy agentic models in accuracy and speed, while also being up to six times cheaper. The r-1 model is currently available in preview and operates through Reducto's hosted Parse API on V3, activated via a configuration flag. Reducto does not offer open weights or local checkpoints for self-hosting; instead, its platform supports multi-tenant cloud, customer VPC, on-premises, and air-gapped installations, with SOC 2 Type II attestation and HIPAA processing available on higher tiers, in accordance with its security policies.

The core innovation of r-1 lies in its consolidation of multiple processing stages into a single pass. Reducto's legacy parsing system involved separate stages for OCR, layout detection, and post-processing, with optional agentic vision-language passes layered on top. Each additional model call in this pipeline introduced latency. In contrast, r-1 integrates the processing of text, tables, figures, layout, reading order, formatting, and grounding into one comprehensive pass. Each identified block of content is returned with page-relative bounding boxes, precisely linking the content to its location on the document. This consolidation addresses a significant pain point for teams working with complex documents such as financial statements, insurance claims, or contracts, who often route files across multiple providers and implement extensive post-processing to achieve acceptable accuracy. Reducto's r-1 model aims to reduce this orchestration cost, rather than solely focusing on raw character accuracy.

In terms of performance metrics, Reducto reported a 20% reduction in error rates for the early preview of r-1 when compared against its own legacy agentic pipelines. The company also stated that internal evaluations showed r-1 outperforming commonly used hyperscaler products and large language models on complex documents. Specifically, Amazon Textract and Azure Document Intelligence were identified in the release as the incumbent baseline category against which r-1 was benchmarked. Regarding pricing, Reducto's legacy agentic models incurred costs ranging from 3 to 6 cents per page, depending on the specific workload. The new r-1 model is priced at 1 cent per page, inclusive of all features, without additional multipliers or credit costs for achieving high accuracy. This significant cost reduction is a key selling point for the new parsing model.