Reducto launches r-1: single-pass parsing model cuts errors 20% at one cent per page

Reducto launched r-1 last week, the first in a new family of parsing models built on a rewritten architecture. The model replaces the company's multi-step agentic OCR pipeline with a single full-page pass and, according to Reducto, is more accurate than its strongest agentic models, faster, and up to six times cheaper.

The legacy Parse pipeline ran OCR, layout detection, and post-processing as separate stages, with optional agentic vision-language model passes layered on top. Each additional model call added latency. r-1 folds text, tables, figures, layout, reading order, styling, and grounding — linking every block to its relative page coordinates — into one full pass. That consolidation is the real product claim: teams working on financial statements, insurance claims, or contracts currently route files across multiple vendors and stitch together post-processing to reach usable accuracy. r-1 targets that orchestration cost, not just raw character accuracy.

Reducto reports a 20% reduction in error rate for the early-preview r-1 versus its previous agentic pipeline. The measurement is against the internal pipeline, not a third-party benchmark. The company also says r-1 outperformed common hyperscaler products and large LLMs on complex documents in internal evaluations, naming Amazon Textract and Azure Document Intelligence as the baseline category. That comparison was vendor-run, with no public benchmark or dataset released alongside the announcement. On pricing, the old agentic models ranged from 3 to 6 cents per page (roughly 11 to 22 agorot) depending on load; r-1 is priced at 1 cent per page (roughly 3.7 agorot), all-in, with no feature multipliers or extra credit costs to reach high accuracy. Reducto frames this as part of a shift to flat product-pricing cards.

According to the documentation, r-1 handles in a single page pass: digital text, scans, and handwriting; table structure with surrounding page context, including merged cells and nested headers; columns, headers, footers, sidebars, and reading order resolved together; figure detection with a generated caption; meaningful styling — headings, lists, bold, underline, strikethrough; and grounding via relative bounding boxes. The long-tail edge cases Reducto flags are dense tables, unusual layouts, low-quality scans, watermarked content, and documents that do not follow an expected template. A missing strikethrough can flip a contract clause, and a misread table can feed an agent the wrong figure, so these edge cases carry weight in regulatory pipelines.

r-1 requires API V3. A Parse request that omits the settings.model parameter still runs on Legacy Parse, so nothing breaks silently; new pipelines created in Reducto's Studio default to r-1. Agentic processing has not disappeared: flows that require custom prompts or advanced graph extraction still route the relevant pages through the agentic pipeline on top of the r-1 output, adding latency. Teams migrating an existing configuration should check the r-1 config compatibility page first, because several legacy settings are ignored or unsupported. Reducto also signaled…