LlamaIndex opens OpenDocRouter, a pay-as-you-go API for swapping document parsers

What did LlamaIndex launch?

Markewich’s 7 October post frames the problem as a flood of OCR and document models: thousands of Hugging Face cards, plus frontier labs shipping readers that are expensive to prompt, host and compare. OpenDocRouter is LlamaIndex’s way of sharing the recipes it already maintains. The product site’s one-line pitch is “any model, your docs.”

The documented endpoint is POST https://www.opendocrouter.ai/v1/parse. The body needs a model ID and a document. The document can be a public HTTPS URL, an upload_id from POST /v1/uploads, or inline base64 up to about 3 MB. Supported types are PDF, PNG and JPEG. URLs and uploads are capped at 50 MB or 500 pages.

That is a parsing API, not a retrieval stack. It emits Markdown, optional layout boxes, and usage. It does not index the pages, answer questions, or claim to replace OCR versus document AI as a category. We did not send it a file.

How does a parse request work?

Each page is a separate model call with its own status and charge. A default sync request, for up to 50 pages, returns HTTP 200 with every page. Status completed means every page worked, partial means some did, and failed means none did. The docs tell you to join pages[].markdown and to check for errors before using the text. Over 50 pages, a sync request is refused as too_large; you send mode async with cache true, poll GET /v1/parse/{id}, then add expand=markdown.

The official clients are pip install opendocrouter and npm install @llamaindex/opendocrouter. PyPI lists version 1.0.0 uploaded on 7 October 2026 at 02:48 UTC, MIT-licensed, Python 3.9 and later. Methods follow the HTTP API: parse.create, parse.get, parse.delete, uploads.create, credits.get and models.list. In Python, the response field model_version is api_model_version because Pydantic reserves the model_ prefix.

Without cache, Markdown exists only in the POST response. If the connection drops, the request can still be charged and a retry is a new request. With cache true, results are stored encrypted for 24 hours, a later identical request is free, and DELETE /v1/parse/{id} removes the stored text sooner. The docs say the document itself is not kept unless you opt into cache, and that an async request’s file is kept only until parsing ends.

Which models are listed, and what do the scores say?

The launch post named ten models. The models page we opened on 8 October lists eleven, with prices “as of version 2026-10-06.” The extra row is anthropic/claude-haiku-5-5, recipe version 2026-10-07, at $0.10 input and $0.50 output per million tokens. That is the same list price Anthropic posted for Haiku 5.5. The other frontier IDs are Claude Opus 5.5, Gemini 3 Flash, Gemini 3.8 Flash Low, GPT-5.6 Terra and GPT-6 Luna. The open-source IDs are Infinity-Parser2-Flash, MinerU2.5-Pro, TeleOCR, dots.mocr and PaddleOCR-VL-1.6.

Overall is the mean of tables, charts, faithfulness, formatting and grounding on LlamaIndex’s catalog. The April 2026 ParseBench paper, by LlamaIndex authors, describes about 2,000 human-verified enterprise pages and more than 169,000 test rules. In that paper LlamaParse Agentic led 14 methods at 84.88 percent. OpenDocRouter’s table is a later vendor catalog, not an independent 8 October leaderboard. Read it the way you would any benchmark-marketing claim: useful for routing, not a ranking we verified.

What does layout grounding add?

Set layout true and each successful page can include a layout object: page size plus elements in reading order. The documented classes are title, section_header, text, list_item, table, picture, chart, formula, caption, footnote, page_header, page_footer, code, form and key_value. Boxes are fractions of the page from the top left. An element that continues across columns can have more than one box. A picture that the Markdown never mentions has lines null.

The launch post says some models emit boxes natively, some need prompting, and some cannot do it at all, so OpenDocRouter applies its own grounding engine. Layout adds $0.20 per million tokens on pages whose layout comes back. If layout fails, the page keeps its Markdown and layout is not charged. That is a convenience layer, not a claim that every model now understands the page the same way.

Grounded boxes help a later RAG pipeline cite a region instead of a blob of Markdown. They do not make the transcription true. A cheap model can still invent a cell or drop a footnote. The docs list page errors such as content_filtered, output_truncated, repetitive_output and invalid_output for exactly those cases.

How do you access it, and what does it cost?

You create an account, take an API key, and set OPEN_DOC_ROUTER_API_KEY. The docs say new accounts start with $5 of free credit and that you top up from $10 in the dashboard, plus a 5 percent fee on each top-up. The 7 October blog said top-ups start from $25; the docs page we opened on 8 October says $10. We are reporting both, and using the docs for the live product.

Frontier models are charged at provider token prices with no markup, according to the docs. Each successful page is charged for its tokens, plus layout when you asked for it. A request must have enough credit for its maximum charge — the model’s most per page times the page count — which is held and then released. Concurrency is 10 requests and 5 async requests at once. POST /v1/parse is limited to 300 calls a minute; polling is 60. Pages still running after 270 seconds come back as timeout. Async jobs wait up to 30 minutes for capacity.

Third-party models receive the pages you send them. The terms addendum we opened says Anthropic, Google and OpenAI usage policies apply to that content, that hosted models run on LlamaIndex’s infrastructure providers, and that layout images go to its OCR and layout services, which “keep no data.” Output is generated and may be wrong.

How is this different from LlamaParse, and what is still unproven?

Markewich’s post is blunt: OpenDocRouter is for hosting the latest models and switching between them while paying for what you use. LlamaParse remains the managed platform, with hand-tuned tiers, enterprise controls, self-hosted deployments, and APIs for schema extraction and indexing. If you already standardized on LlamaParse, this launch does not retire it.

We have not parsed a document, compared recipes, or checked whether a given model ID is self-hosted or proxied. ParseBench categories are LlamaIndex’s. Cost per 1,000 pages is an estimate from tokens used on that corpus, and the catalog warns your pages may use more. A public HTTPS URL on the default port is required for the URL form; encrypted PDFs are unreadable_document. That is enough to try a parser. It is not a reason to delete your current pipeline on launch day.