Choosing a model to read your documents has turned into a job of its own, and LlamaIndex says so in the first paragraph of the post announcing OpenDocRouter on 7 October. Thousands of OCR models sit on HuggingFace, the company writes, and frontier labs ship new ones that also read documents well, sometimes at steep prices. Every team that uses them repeats the same chores: tuning prompts, handling rate limits, paying for hosting, and benchmarking each newcomer.
OpenDocRouter is LlamaIndex’s answer. It is a hosted API that takes a document and returns Markdown, with the developer picking which model does the reading. The launch lineup has ten: five frontier models (Claude Opus 5.5, two Gemini Flash versions, GPT-5.6 Terra, and GPT-6 Luna) and five open-source ones, including MinerU2.5-Pro and PaddleOCR-VL-1.6.
The limits are what a developer will want to see first. The service accepts PDFs, PNG and JPEG files, or links to any of those. A synchronous call, where the result comes straight back, works for documents up to 50 pages. Anything bigger has to run as an asynchronous job, which means your code polls until it finishes. A single request tops out at 500 pages or 50MB.
LlamaIndex describes the behavior of each model as a “versioned recipe,” its term for the bundle of prompt, processing steps, and settings wrapped around the model. The idea is practical. If you pin a recipe, the same document should keep producing the same kind of output even after the underlying model is updated or the router’s operator changes its prompt. Anyone who has watched a parsing pipeline quietly shift after a model refresh will see the appeal, though the post does not say how long old recipe versions stay available.
Models disagree about bounding boxes, the coordinates that say where each piece of text sat on the page. Some produce them natively, some need coaxing, and some cannot do it at all. LlamaIndex built what it calls a grounding engine that adds boxes and reading-order layout to any model’s output, with a fixed set of fifteen element types such as tables, formulas, and footnotes. It costs an extra $0.20 per million tokens.
Pricing is pay as you go, with credits starting at a $25 top-up, and pages that fail during processing are not billed. The spread between models is the part worth studying. By LlamaIndex’s own ParseBench measurements, a thousand pages costs $48.82 on Claude Opus 5.5, $0.80 on GPT-6 Luna, and $0.86 on MinerU2.5-Pro. ParseBench is LlamaIndex’s benchmark, and the company sells the service being compared, so those figures are its claim about its own platform and not an independent test. The post also publishes no quality scores in the text that would let a reader weigh cost against accuracy without clicking through to the benchmark pages.
A router is also a middleman. Every document that passes through OpenDocRouter passes through LlamaIndex, which gains data on how models perform and a customer relationship that sits above the model makers. That is the ordinary logic of an aggregator, and the company already sells a pricier managed product, LlamaParse, with hand-tuned tiers and self-hosted options. It positions OpenDocRouter as the lighter alternative for people who want to switch models freely.
The more durable point is that document parsing is now priced like a commodity with a range of roughly sixtyfold. A team still sending every page to a premium model can find out cheaply whether a model costing under a dollar per thousand pages holds up on its own files, and a router makes that test a one-line change.
Based on LlamaIndex’s blog post introducing OpenDocRouter, published 7 October 2026.
