OpenDocRouter: document parsing at $0.80 per 1,000 pages

LlamaIndex launched OpenDocRouter today: one document parsing API that sends your PDFs and images to frontier or open-source models and gives back markdown. The interesting part is not the API. It is the price sheet. The top model costs $48.82 per 1,000 pages. A model that scores 13 points lower costs $0.80. That is a 61x spread for one parsing job, and most teams pay the top price for every page.

What actually happened

OpenDocRouter is a hosted document-to-markdown service, separate from LlamaIndex's managed LlamaParse product. Each model runs a versioned "recipe" of prompts, processing and settings, so output stays stable when you pin a version. Per the announcement, it takes PDF, PNG and JPEG files, returns synchronous results up to 50 pages, and caps a request at 500 pages or 50MB.

The models page lists 11 models with a ParseBench overall score and cost per 1,000 pages. A few rows:

• Claude Opus 5.5: 84.20, $48.82
• Gemini 3 Flash: 79.70, $19.67
• GPT-6 Luna: 71.34, $0.80
• MinerU2.5-Pro (open source): 70.05, $0.86
• Claude Haiku 5.5: 70.01, $1.22

Billing is per token from prepaid credits. The site says frontier models cost exactly what their providers charge, and failed pages are not charged. An optional layout mode returns bounding boxes for $0.20 per million extra tokens.

Why document parsing cost matters for your business

Route by document, not by brand. A clean, typed vendor invoice does not need the best model on the board. A faded, handwritten delivery ticket might. Send the easy 80% to a sub-$1 model and the hard 20% to the expensive one. On 10,000 pages a month, that is the difference between about $490 and about $100.

ParseBench is the vendor's benchmark. LlamaIndex built it. Scores are a starting point, not a verdict. Take 50 of your real documents, parse them with two or three models, and compare against a hand-checked answer key.

Pinned recipes are the quiet feature. Parsing pipelines break when a model update changes table output. Versioned recipes give you a fixed target, so a model swap is a decision you make, not a surprise.

Sources: LlamaIndex, OpenDocRouter models and pricing.