Reducto Releases r-1: A Single Pass Document Parsing Model That Cuts Errors 20% at 1 Cent Per Page

Document parsing sits under every RAG stack, claims pipeline, and contract agent that still fails on merged cells and strikethroughs. Reducto's r-1, now in preview on the V3 Parse API, replaces a staged OCR-plus-layout-plus-agentic-VLM chain with a single full-page pass that returns text, tables, figures, reading order, formatting, and page-relative bounding boxes together.

The commercial pitch is blunt: legacy agentic Parse ran 3–6 cents per page; r-1 is 1 cent flat, with Reducto claiming up to 6x cheaper and a 20% error cut versus its own prior pipelines. That matters less as a marketing score and more as an attempt to kill the multi-vendor orchestration tax finance and insurance teams still pay to reach usable accuracy.

The architectural claim is the real news, not the percentage. Folding OCR, layout, tables, figures, formatting, and grounding into one pass attacks latency and failure modes that accumulate when each stage can disagree with the next. Bounding boxes that tie every block back to the page are the compliance feature: regulators and auditors care where a number lived, not only that a model emitted it.

Treat the accuracy story with the caveat MarkTechPost itself flags. The 20% error reduction is versus Reducto's legacy agentic stack, not a public third-party harness. Head-to-heads against Amazon Textract, Azure Document Intelligence, and large LLMs are vendor-run, with no shared dataset released. That does not make the claim false; it makes it non-comparable until someone else reproduces it on a fixed corpus of dense tables, watermarked scans, and template-free layouts—the long-tail cases Reducto itself calls out.

The migration path is carefully non-breaking: omit `settings.model` and you stay on legacy Parse; Studio defaults new pipelines to r-1. Agentic is not dead—custom prompts and advanced chart extraction still bolt onto the r-1 result and add latency. Closed weights and hosted-only preview keep Reducto in the API-business column, even as the broader platform already sells VPC, on-prem, and air-gapped installs with SOC 2 Type II and HIPAA options.

**Signal Desk's read:** this is a price-and-pipeline consolidation play against hyperscaler document APIs and against the DIY agentic OCR stacks enterprises bolted together when those APIs were not good enough. The flat cent-per-page rate card is more disruptive than the 20% number if it holds under production mix. Watch whether r-1 mini and automatic per-page routing arrive before customers start asking for the eval harness Reducto has not published.

Reducto frames r-1 as the first model in a rewritten parsing family, not a patch on the old agentic OCR. Teams comparing parsers can request up to $5,000 in credits for a side-by-side run—useful only if they bring their own hardest pages, not marketing PDFs.