pdfmux vs Reducto: self-hosted library or hosted agentic API?
Direct answer: pdfmux and Reducto solve overlapping problems from opposite ends. Reducto is a hosted agentic document platform — Parse, Extract, Split, Classify, and Edit APIs, credit-metered, with a Studio UI for building pipelines without code. pdfmux is a self-hosted Python library — pip install pdfmux, seven local extraction backends run for free, with paid vision-LLM fallback (BYOK) only for the hardest pages. Pick Reducto when you want a managed platform with zero-data-retention agreements and a no-code pipeline builder. Pick pdfmux when you want to own the extraction step, avoid per-page billing on the documents that don’t need it, and run inside your own infrastructure.
What Reducto actually is
Reducto is an agentic document platform built around six operations: Parse (documents into structured JSON), Extract (typed JSON with citations and bounding boxes), Split (segment multi-document files), Classify (route documents via plain-language taxonomies), Edit (fill extracted data back into PDF/DOCX), and Studio (a no-code interface for building and testing pipelines). It supports 30+ file types including scans, handwriting, charts, and spreadsheets, and ships Python/Node SDKs, a CLI, and both local and hosted MCP servers.
Reducto’s pricing is credit-based:
| Plan | Price | Throughput | Notes |
|---|---|---|---|
| Standard (pay-as-you-go) | Free up to 15,000 credits, then $0.015/credit | 200 concurrent pages | Up to 5 Studio seats |
| Growth | Custom | 350 concurrent pages | Volume discounts, zero-data-retention, BAA, EU/AU residency |
| Enterprise | Custom | 500+ concurrent pages | VPC/on-prem, custom MSA/SLA, SSO/SAML |
Credit consumption depends on operation type and configuration rather than a flat per-page rate, and async batch jobs get a 20% credit discount. This is the standard model for a hosted document AI vendor: you’re paying for infrastructure, model orchestration, and support SLAs on every document, regardless of how simple that document is.
What pdfmux actually is
pdfmux is an open-source Python library (pip install pdfmux, ~15 MB) that routes each page to the cheapest backend that can handle it, rather than sending everything through one model:
| Backend | Handles | Cost |
|---|---|---|
| PyMuPDF | Digital text | $0 |
| RapidOCR | Scanned, CPU OCR | $0 |
| Surya | Heavy OCR fallback (GPU) | $0 |
| Docling | Table-heavy documents | $0 |
| OpenDataLoader | Complex multi-column reading order | $0 |
| Mistral OCR (cloud) | Cloud table OCR | $0.002/page |
| Vision LLM (BYOK — Gemini/Claude/GPT-4o) | Charts, handwriting, hardest pages | ~$0.005–$0.015/page |
The classifier decides per page which backend to use, then an audit step scores confidence and re-routes low-confidence pages to the next backend in the chain. Most digital and scanned PDFs never touch a paid API at all — you only pay when a page genuinely needs a vision LLM.
Feature comparison
| Feature | pdfmux | Reducto |
|---|---|---|
| Deployment | Self-hosted (your infra) | Hosted API + Studio UI |
| Pricing model | Free local backends, pay-per-page only on BYOK fallback | Credit-metered on every page |
| No-code pipeline builder | No | Yes (Studio) |
| Document splitting/classification | Not built in | Native (Split, Classify APIs) |
| Bounding boxes / citations | Confidence + warnings per page | Citations + bounding boxes on Extract |
| Data residency control | Full (runs where you run it) | EU/AU on Growth+, VPC/on-prem on Enterprise |
| Open source | Yes (MIT) | No |
| Setup | pip install pdfmux | API key + SDK/CLI/MCP |
| Install footprint | ~15 MB (base) | N/A (hosted) |
When Reducto wins
- You want a no-code pipeline builder. Studio lets non-engineers configure and test extraction pipelines without touching an API.
- You need document splitting and classification as first-class operations, not just extraction — Reducto’s Split and Classify APIs handle multi-document files and routing natively.
- You want a vendor-managed compliance posture. Zero-data-retention agreements, BAA, and SOC-style enterprise controls are turnkey on Growth/Enterprise plans instead of something you build yourself.
- You don’t want to run or maintain extraction infrastructure at all — Reducto is fully hosted; there’s nothing to deploy, patch, or scale.
When pdfmux wins
- Cost predictability at volume. Reducto meters credits on every page regardless of complexity. pdfmux runs the majority of pages through free local backends and only bills for the vision-LLM fallback on genuinely hard pages — at high volume with mostly-digital or mostly-scanned documents, this is a materially different cost curve.
- You need it to run inside your own infrastructure — air-gapped environments, strict data-residency requirements you control directly, or CI/batch pipelines where an external API call per page isn’t acceptable.
- You’re already deep in a Python pipeline (LangChain, LlamaIndex, custom RAG) and want extraction as a library call, not a separate hosted service to integrate and monitor.
- You want to inspect and swap the extraction backend yourself — pdfmux is MIT-licensed and each backend (Docling, Surya, RapidOCR) is a normal open-source dependency you can pin, audit, or replace.
Cost at scale: a worked example
Cost comparisons between a credit-metered API and a mostly-free library depend entirely on your document mix, so treat the numbers below as an illustrative model, not a quoted benchmark — plug in your own page counts and mix to get a real answer.
Say a team processes 100,000 pages a month, split roughly 60% clean digital PDFs, 30% clean scans, and 10% hard cases (handwriting, low-quality scans, dense charts):
- On Reducto, every page is credit-metered regardless of which of the three buckets it falls into. At $0.015/credit and even a conservative 1 credit/page for the simplest operation, that’s a cost floor of roughly $1,500/month before volume discounts — and complex operations (Extract with citations, Split, Classify) consume more credits per page.
- On pdfmux, the 60% digital pages route to PyMuPDF for $0, the 30% clean scans route to RapidOCR for $0, and only the 10% hard cases (10,000 pages) hit the BYOK vision-LLM fallback at roughly $0.005–$0.015/page — $50 to $150/month total.
The gap isn’t because one vendor is “better” — it’s architectural. Reducto prices for the hardest page in every request because its model has to be ready to handle it. pdfmux prices per page because routing happens first, so easy pages never touch a paid model at all. If your document mix is mostly hard cases (dense charts, handwriting-heavy forms), this gap narrows or reverses, since you’d be paying the BYOK rate on most pages anyway — at that point Reducto’s managed infrastructure and citation tooling may be worth the premium.
Where the two overlap least
Reducto’s Split and Classify operations solve a problem pdfmux doesn’t attempt: given a folder of mixed, unlabeled multi-document PDFs (say, a batch of scanned mail that mixes invoices, contracts, and correspondence), automatically separate and route them. pdfmux assumes you already know what kind of document you’re feeding it and focuses entirely on getting clean text and structure out of it. If document triage is your actual bottleneck — not extraction accuracy — Reducto’s Classify API addresses that directly, and building an equivalent classifier yourself on top of pdfmux is extra work Reducto has already done.
Conversely, pdfmux’s pdfmux estimate and pdfmux doctor commands are built for a self-hosted operator who wants to predict cost and diagnose backend gaps before running a job — tooling that doesn’t have a direct Reducto equivalent, because on a hosted platform the provider owns that operational surface for you.
Quick code comparison
pdfmux:
import pdfmux
result = pdfmux.process("contract.pdf", quality="standard")
print(result.text)
print(result.confidence) # e.g. 0.91Reducto (Python SDK):
from reducto import Reducto
client = Reducto(api_key="...")
result = client.parse.run(document_url="https://.../contract.pdf")
print(result.result)Both give you structured output from an unstructured document. The difference is where the compute runs and who you pay for it — pdfmux ships that decision to you, Reducto ships it to Reducto’s infrastructure.
FAQ
Can I self-host Reducto?
Only on the Enterprise plan, which adds VPC and on-premises deployment with a custom MSA. Standard and Growth are fully hosted.
Does pdfmux have a no-code UI like Reducto Studio?
No. pdfmux is a library and CLI (pdfmux convert, pdfmux serve for MCP). If you want a pipeline builder without writing code, Reducto’s Studio is purpose-built for that.
Which has better accuracy?
Neither vendor publishes a shared, independently-run benchmark against the other, so any specific accuracy number here would be a guess dressed up as a fact — we don’t publish those. If accuracy on your specific document set is the deciding factor, the only reliable answer is to run both against a sample of your own documents and compare.
Can I use both?
Yes — some teams route the bulk of routine documents (invoices, forms, digital PDFs) through pdfmux for cost reasons and reserve a hosted platform like Reducto for complex, high-stakes documents where citations and bounding boxes matter for audit trails.
For a broader look at the self-hosted landscape, see pdfmux vs Docling and pdfmux vs Unstructured. For hosted-API alternatives, see pdfmux vs LlamaParse.