Reducto r-1 Cuts Document Parsing Cost to 1¢ Per Page With Single-Pass Model

Reducto has shipped r-1, the first model in a new parsing family that replaces the company's multi-stage agentic OCR pipeline with a single full-page inference pass. The architectural consolidation collapses what previously ran as discrete OCR, layout detection, and post-processing stages — with optional vision-language model calls layered on top — into one operation that jointly resolves text, tables, figures, reading order, formatting, and page-relative bounding boxes. For engineers maintaining document ingestion pipelines across financial statements, insurance claims, or contracts, the claim is straightforward: the same or better accuracy at a fraction of the orchestration cost, available today in preview on the V3 Parse API.

The launch is a concrete test of whether pipeline architecture, not better models, drives AI gains. Reducto collapsed four sequential model calls into one pass and reports measurable accuracy and cost improvements as a result.

Architecture: One Pass, Everything Grounded

Legacy Parse operated as a sequential system: OCR extracted text, a layout model detected regions, optional agentic vision-language passes handled ambiguous content, and a final merge step assembled outputs. Each additional model call added latency and introduced another surface for error propagation.

r-1 processes the full page in one shot, handling: digital text, scans, and handwriting; table structure including merged cells and nested headers read in context of surrounding page content; column and multi-region layout with reading order resolved simultaneously; figure detection with a short generated description; meaning-bearing formatting such as headings, lists, bold, underlines, and strikethroughs; and page-relative bounding boxes attached to every output block. A dropped strikethrough that inverts a contract clause or a misread merged cell that passes a downstream agent the wrong figure are the failure modes Reducto specifically targets.

Agentic processing is not eliminated. Workflows requiring custom prompts or advanced chart extraction still route those pages through the agentic pipeline, layered onto the r-1 result — with the associated latency cost. Teams migrating existing configurations should consult Reducto's r-1 compatibility documentation, as some legacy settings are ignored or unsupported.

Benchmarks and Caveats

Reducto reports a 20% reduction in error rate for the r-1 preview compared with its own legacy agentic pipelines. Internal evaluations also show r-1 outperforming Amazon Textract, Azure Document Intelligence, and large LLMs on complex documents. The 20% error reduction is measured against Reducto's own prior pipeline, not an independent third-party baseline, and the head-to-head against hyperscalers is vendor-run with no public evaluation harness or dataset released alongside the announcement. Developers should treat these figures as directional until independent benchmarks emerge — a recurring challenge in AI capability verification.

Pricing and Competitive Position

Legacy agentic models ran at 3 to 6 cents per page depending on workload. r-1 is priced at a flat 1 cent per page, all in, with no feature multipliers or credit costs added to reach high-accuracy outputs — up to 6x cheaper at the high end.

Per-page figures are straight arithmetic on Reducto's published rates and exclude Extract, Split, Classify, and Edit usage, as well as any agentic augmentation added on top of r-1. Reducto frames this pricing as part of a wider move to flat product rate cards across its platform.

Availability and Roadmap

r-1 is live in preview via Reducto's hosted Parse API on V3, activated with a single configuration flag: settings={"model": "r-1"}. A Parse request that omits settings.model continues to run on legacy Parse, so existing integrations do not break silently. New pipelines created in Studio default to r-1. There are no open weights and no self-hostable checkpoint; deployment runs through Reducto's hosted infrastructure, which supports multi-tenant cloud, customer VPC, on-premises, and air-gapped installs, with SOC 2 Type II attestation and HIPAA processing on higher tiers.

Two follow-on releases are signalled: r-1 mini, targeting speed- and cost-sensitive workloads, and automatic per-page routing that selects the appropriate model for each page within a document. Organisations migrating from another parser can request up to $5,000 in credits for a side-by-side comparison.

Collapsing four sequential model calls into one pass delivers both a 20% reported accuracy gain and a cost reduction of 50 to 83% without requiring larger models or more compute. Whether independent benchmarks validate the accuracy claims will determine how quickly r-1 displaces Textract and Azure Document Intelligence in regulated pipelines where a misread table or a missed strikethrough carries legal exposure, not just latency. That tension between architectural specificity and GPU scaling is precisely what this release puts to a real-world test.