Reducto

What it is

Reducto is a document parsing and extraction API for teams whose AI workflow keeps failing on the document rather than on the model. Adit Abraham and Raunak Chowdhuri founded it out of Y Combinator after building long-term memory systems for LLMs and concluding the unsolved piece was turning messy PDFs into structured data an agent can act on. Reducto reads layout, format and structure as signals of meaning instead of flattening a page into a text blob — a redline in a contract, a handwritten dosage, a table running across three pages of a filing.

The product is six endpoints, not a chat window. Parse converts a document to structured JSON with layout-aware chunking, Extract pulls named fields into a schema you define, Split segments by natural-language rules, Classify routes by document type, Pipelines chains the steps — and Edit writes changes back into PDF and DOCX, which Reducto calls one of the first document-editing APIs and is the part competitors do not have. Underneath is r-1, a single-pass full-page model that handles text, tables, figures and formatting together instead of stitching OCR output onto a separate layout detector. There are Python, Node and Go SDKs, a REST API, a CLI, an MCP server, and Studio, a visual workspace for testing a pipeline before you wire it into code.

a16z led a $75M Series B announced on 14 October 2025, taking total funding to $108M with Benchmark, First Round Capital, BoxGroup and Y Combinator participating, six months after a $24.5M Series A. Reducto reported more than 1 billion pages processed and a 6x rise in monthly volume between the two rounds, and names Harvey, Mercor and Rogo as customers. It made its first acquisition on 7 May 2026, absorbing the team behind AI study notebook Opennote.

Why it shows up in ops stacks

Because the bottleneck sits upstream of the prompt. A contract-review assistant that misses a struck-through clause, an invoice pipeline that reads a merged cell into the wrong column, an ATS that mangles a two-column CV — none of that is fixed by a better model, and every hour spent tuning the prompt is spent on the wrong layer. Reducto is the argument that extraction should be bought rather than maintained.

For legal-ops that is the document layer under Ironclad or Spellbook, where the source material is scanned executed copies rather than clean templates. For RevOps it is order forms, signed MSAs and pricing exhibits that never reach the CRM. For recruiting it is resume parsing, the weakest link in every ATS for a decade.

Deployment is the other reason it clears security review: SaaS, hybrid VPC, full VPC or air-gapped on-prem, with SOC 2 Type II, a HIPAA BAA, AES-256 at rest, TLS 1.2+ in transit, EU data residency and 24-hour document deletion.

A use case worth buying it for

The contract-extraction job that stalls in-house legal teams: several thousand executed agreements, half of them scans, and a need to pull governing law, renewal dates, assignment terms and liability caps into an auditable table. Classify routes by agreement type, Split isolates exhibits, Parse produces layout-aware JSON, Extract fills the schema. Every field traces back to a page coordinate — which is what makes the output defensible where a general model summarizing a PDF is not.

Pricing

Pricing is per endpoint and per page, effective 1 September 2026, with no seat licence. Parse runs $10-$30 per 1,000 pages by complexity, Extract and Split $20-$40, Classify $7.50-$15, and Edit $60 per 1,000 pages or $15 per 1,000 pre-filled. Spreadsheets meter by cell, and batch jobs take a 20% discount. Standard starts with $150 of free usage — 15,000 pages of list-rate Parse.

The headline cent-a-page figure is the floor, not the bill. A parse-only pilot at 100,000 pages a month is $1,000 a month. A four-step production pipeline — classify, split, parse, extract — bills all four meters against the same page, so 250,000 pages a month lands near $14,400 a month at list, or roughly $173K a year before volume discounts. That is the band to budget against.

Growth and Enterprise are quote-only. Growth adds volume discounts, data residency and unlimited Studio seats; Enterprise adds VPC and on-prem, custom SLA, SSO/SAML and RBAC. Concurrency is the other tier gate — 200 pages on Standard, 350 on Growth, 500+ on Enterprise. A startup programme gives free credits and discounted Growth rates to teams under $15M raised, $3M revenue and 50 employees.

Best for

Legal-ops and finance-ops leaders with an AI workflow already built, already plateaued on extraction quality, burning engineering time prompt-tuning a general model against documents it was never going to read correctly. The economics work above roughly 50,000 pages a month, on documents that are genuinely hard — scans, dense tables, multi-column layouts, handwriting.

Do not buy it for clean, uniform, machine-generated forms — if every document is the same digital invoice template, a hyperscaler OCR service does the job for a fraction of the cost. Do not buy it as an end-user application either. Reducto ships APIs and a builder workspace; someone on your team writes the integration. If nobody will own that code, buy a workflow product with extraction inside it.

Alternatives, and when to pick them instead

• Azure AI Document Intelligence — the incumbent by installed base, and the default when the company already sits on an Azure commitment. Pick it when documents are standard forms, invoices and IDs, and procurement counts more than table fidelity. Reducto’s own RD-TableBench put Azure at 82.7% against Reducto’s 90.2% on 1,000 complex tables; that benchmark is vendor-run, so treat the ranking as a hypothesis to test.
• AWS Textract — the other incumbent, and the right pick for an AWS-native pipeline feeding Bedrock. Note that Textract is not the cheap option it is assumed to be: forms-and-tables extraction runs about $0.065 a page, which is above Reducto’s list Parse rate. It scored 80.9% on the same vendor-run table benchmark.
• Mistral OCR — the fastest-growing entrant, and the price floor. OCR 4 shipped on 23 June 2026 at $4 per 1,000 pages ($2 in batch) with paragraph-level bounding boxes, confidence scores and 170 languages; OCR 3 stays at $2 per 1,000 ($1 batch). Pick it when you need bulk text and layout at the lowest unit cost and complex-table fidelity is not the binding constraint. At 1M pages a month the gap against a full Reducto pipeline is five figures monthly.

Watch-outs

• The accuracy gap you are buying is vendor-measured. RD-TableBench is Reducto’s own benchmark on Reducto’s own table set. Guard: before signing, run a 200-document eval on your ugliest real files against hand-labelled ground truth, scoring the same fields you will extract in production. The $150 free tier covers 15,000 Parse pages, which is more than enough — and if the lead does not survive your own documents, you have learned the thing that matters.
• Four endpoints on one page means four meters. The cent-a-page Parse rate is the number that gets quoted internally; a classify-split-parse-extract pipeline is closer to $0.058 a page, a 5.8x difference on the same document. Guard: cost the full pipeline per page before the pilot, set dollar-threshold alerts in Studio, and pull the Usage Export API into whatever tracks your cloud spend.
• Everything security review asks for sits on Enterprise. SSO/SAML, RBAC, custom SLA, VPC and on-prem are all above the Growth line, and zero-retention terms plus the HIPAA BAA are Growth-and-up. Guard: if privileged or regulated material is in scope, price Enterprise from the first call and get the retention terms in the contract — the Standard tier is a prototyping tier, not the one you send client documents through.
• A two-year-old company now holds a layer your production workflow depends on. Reducto is an acquirer rather than a target, which lowers the near-term risk, but the concentration is real. Guard: keep the extraction call behind your own interface so a swap to Azure, Textract or Mistral is a config change, and archive the raw source documents, not only the parsed JSON.