Reducto’s parsing model cuts document ingestion errors by nine percentage points

Abraham highlights that even high-accuracy parsers can fail catastrophically if they miss small but critical details. Reducto’s latest model, r-1, combines Agentic OCR (vision-language models) with formatting and layout detection to achieve 9% higher macro accuracy in enterprise tasks while reducing reasoning tokens by 13% compared to frontier models processing raw documents. Customers like Harvey and Scale AI report fewer OCR-related errors and improved RAG/search performance after switching to Reducto. The company now processes billions of pages monthly, optimizing for cost (one cent per page), latency, and scalability—balancing efficiency with accuracy for complex documents.