Last week, Reducto announced r-1. It is the first model in a new parsing family built on a rewritten architecture, and it replaces the company’s multi stage agentic OCR with one full page pass. Reducto says r-1 is more accurate than its most powerful legacy agentic models, faster, and up to 6x cheaper.
Is it deployable today? Yes, in preview. r-1 runs through Reducto’s hosted Parse API on V3 and is switched on with a configuration flag. There are no open weights and no local checkpoint to self host. Reducto’s platform separately supports multi tenant cloud, customer VPC, on premises, and air gapped installs, with SOC 2 Type II attestation and HIPAA processing on higher tiers, per its security policies.
One pass instead of a pipeline
Legacy Parse runs OCR, layout detection, and post processing as separate stages, with optional agentic vision language passes layered on top. Each extra model call adds latency. r-1 folds text, tables, figures, layout, reading order, formatting, and grounding into a single full page pass. Every block returns with page relative bounding boxes that tie content back to its position on the page.
The consolidation is the real product claim. Teams working on financial statements, insurance claims, or contracts often route files across several providers and bolt on post processing to reach usable accuracy. r-1 targets that orchestration cost, not only raw character accuracy.
The numbers
Reducto reports a 20% reduction in error rate for the early r-1 preview against its own legacy agentic pipelines. It also states that r-1 outperformed commonly used hyperscaler products and large LLMs on complex documents in internal evaluations. Amazon Textract and Azure Document Intelligence are named in the release as the incumbent baseline category.
On price, legacy agentic models ran between 3 and 6 cents per page depending on workload. r-1 is 1 cent per page, all in, with no feature multipliers or credit costs layered on to reach high accuracy. Reducto frames this as part of a wider move to flat product rate cards.
Please note that the 20% error reduction is measured relative to Reducto’s own prior pipeline, not to a third party baseline. And the head to head against hyperscalers and LLMs is vendor run, with no public eval harness or dataset released alongside the announcement.
What r-1 resolves on the page
Per the documentation, r-1 handles the following natively in the full page pass: digital text, scans, and handwriting; table structure read with surrounding page context, including merged cells and nested headers; columns, headers, footers, sidebars, and reading order resolved together; figure detection with a short generated description; formatting that carries meaning, including headings, lists, bold, underlines, and strikethroughs; and grounding through page relative bounding boxes.
The long tail cases Reducto calls out are dense tables, unusual layouts, low quality scans, watermarked content, and documents that follow no predictable template. A dropped strikethrough can invert a contract clause, and a misread table can hand an agent the wrong figure, which is why these edge cases carry weight in regulated pipelines.
