Reducto has officially launched r-1, a new parsing model that ditches the traditional multi-stage pipeline for a unified, single-pass architecture. Instead of chaining separate calls for OCR, layout detection, and post-processing, r-1 processes text, tables, figures, and formatting in one go. The company claims this shift delivers a 20% reduction in error rates compared to its legacy agentic models while slashing latency. For engineering teams tired of orchestrating complex workflows across multiple providers, this consolidation targets the hidden tax of integration rather than just raw character accuracy.
The pricing strategy is aggressive. Legacy agentic models cost between 3 and 6 cents per page, depending on workload complexity. r-1 drops that to a flat 1 cent per page, all-in, with no hidden feature multipliers. Reducto positions this as a direct challenge to hyperscaler incumbents like Amazon Textract and Azure Document Intelligence, stating that internal evaluations show r-1 outperforming these baselines on complex documents. However, it is crucial to note that the reported 20% error reduction is measured against Reducto’s own prior pipeline, not an independent third-party benchmark. No public evaluation harness or dataset was released alongside the announcement, leaving some room for skepticism regarding head-to-head performance against external competitors.
Technically, the model handles dense tables, merged cells, and even meaningful formatting like strikethroughs natively. Each block returns page-relative bounding boxes, ensuring content remains grounded in its original position. This matters because a missed strikethrough can invert a contract clause, and a misread table can feed bad data into downstream agents. While r-1 runs through the hosted V3 Parse API via a configuration flag, there are no open weights or local checkpoints available for self-hosting. Enterprise users can still deploy via customer VPC or air-gapped installs with SOC 2 Type II and HIPAA compliance on higher tiers.
Migration requires opting in via settings.model: "r-1" in the API call; legacy requests remain unchanged. Agentic processing isn’t dead, though. Workflows needing custom prompts or advanced chart extraction will still route through the older pipeline, adding latency back into the mix. Reducto hints at future releases, including an r-1 mini for speed-sensitive tasks and automatic per-page routing. To ease the transition, organizations moving from other parsers can request up to $5,000 in credits for side-by-side comparisons. The real test will be whether the single-pass approach holds up against the chaotic variety of real-world scanned documents without the safety net of agentic correction.
