LlamaIndex OCRia-1: The 4B Document Parser That Topped ParseBench — and What a Page Really Costs

Most of the information a business actually runs on is trapped in documents that were never designed to be read by software. Financial statements, insurance filings, tax forms, contracts, research papers: the numbers that matter sit inside tables, charts, footnotes and scanned pages, and the tools that digitize them have always traded accuracy for cost or cost for accuracy. LlamaIndex, the company behind the LlamaParse document service, says its new model breaks that trade-off. OCRia-1, introduced on October 22, 2026, is a 4-billion-parameter vision-language model built for one job: turning messy documents into clean, structured data an LLM can actually use.

In short: OCRia-1 scores 87.4 on LlamaIndex's own ParseBench, 12.1 points ahead of the next-highest parsing model, and it now powers LlamaParse at the same price as before. Its strongest results are on tables and charts, the two document elements where most parsers fall apart. The price, though, is where the marketing and the arithmetic part ways: LlamaIndex advertises under half a cent a page, but the current credit table works out to about a cent a page on the cheapest tier, and three cents on the agentic tier most teams will actually use.

What OCRia-1 is

OCRia-1 is not a general-purpose chatbot or a vision model that happens to read documents. It is a specialized 4-billion-parameter vision-language model (VLM) trained specifically for document parsing, and it is the engine now running inside LlamaParse, LlamaIndex's commercial parsing service. Feed it a PDF, a scanned image, a spreadsheet or a slide deck and it returns structured output: text with its layout preserved, tables rebuilt as tables, charts converted into data, mathematical formulas transcribed into LaTeX, and formatting marks like superscripts, subscripts and checkboxes retained in Markdown and HTML. LlamaIndex's launch post describes it as the company's most capable document parsing model.

LlamaIndex built the model on a training set of more than 100,000 document pages, chosen for the cases where current vision models still fail: dense tables, charts that span pages, mathematical notation, multilingual text and difficult scans. Ground-truth annotations came from a mix of generation-and-review vision models and human annotators, with synthetic document generators filling in the long tail of layouts real documents rarely produce. The post-training recipe pairs supervised fine-tuning with reinforcement learning using verifiable rewards, a combination aimed at generalization across table and chart structures rather than memorization of one document type.

The company describes its parsing pipeline as agentic: task-specific agents route different parts of a document — text, charts, tables, handwriting — to the right expert, then recursive checks detect and correct errors. That architecture matters because a parser's job is not transcription. A transcription that gets every digit right but puts a revenue figure under the wrong column header is worse than useless in a financial workflow; it is wrong with confidence. OCRia-1's pitch is that it preserves the relationships between values and their labels, not just the values.

The benchmark numbers, and whose they are

Here is the headline table from LlamaIndex's launch post, measured on ParseBench:

That 87.4 is the highest score among the systems LlamaIndex evaluated, and the 12.1-point gap over the runner-up is large by benchmark standards. But the first thing to note is the provenance: ParseBench is LlamaIndex's own benchmark, released in April 2026 with around 2,000 human-verified pages and more than 167,000 test rules across 14 parsing methods. A vendor topping its own benchmark is expected, not surprising, and the number should be read as a claim to check, not a verdict.

The check exists, and it is less flattering. On olmOCR-bench, an independent benchmark built by the Allen Institute for AI, OCRia-1 scores 83.5 overall against 82.3 for olmOCR-2 — a lead of 1.2 points, not 12. The gap between the two benchmarks is itself informative: olmOCR-2, which looks hopeless on ParseBench at 43.3, is nearly tied with OCRia-1 on its home turf. Different benchmarks reward different things, and the honest reading is that OCRia-1 leads clearly on LlamaIndex's test and narrowly on the independent one. Neither number on its own settles the question.

Tables and charts are the real story

The overall score hides where OCRia-1 actually separates from the field. On ParseBench's tables dimension it scores 91.5, ahead of Chandra-OCR-2 at 89.2 and LightOnOCR-3-4B at 82.5. On the three independent table benchmarks LlamaIndex cites, it leads as well: 92.9 on olmOCR-bench's tables category, 84.1 on PulseBench-Tab's 1,820 multilingual tables, and 83.2 and 81.9 on the two long-table categories of fr-bench-pdf2md.

Charts are the wider gap. On ParseBench's chart evaluation, OCRia-1 scores 85.1, a 19-point lead over the next-closest model at 66.1, while conventional parsers score near zero — dots.mocr manages 0.1, Surya OCR 2 manages 1.2. The reason this matters is that a chart summary is not enough for the workflows that need parsing. "Revenue grew across all segments" cannot feed a spreadsheet; the actual numbers, matched to the right series and labels, can. OCRia-1 converts charts into HTML tables of their data points, and ParseBench only credits a value when it lands under the correct row and column headers. That is a demanding test, and it is the one dimension where OCRia-1's lead looks structural rather than marginal.

Formatting is the third strength: 83.4 on ParseBench's formatting dimension against 67.9 for the next-highest system. Most parsers strip inline styling; OCRia-1 keeps superscripts, subscripts, strikethroughs and form checkboxes, and writes math as LaTeX rather than mangled plain text. For research papers and technical manuals, that is the difference between a usable parse and a cleanup job.

Speed, the third leg of LlamaIndex's claim, is the least documented. The company reports about 4,700 tokens per second on a single NVIDIA B200, a figure quoted from founder Jerry Liu's launch thread and noted in independent launch coverage. No independent latency test exists yet, and no latency or throughput guarantee is published on the pricing pages. Treat the speed number as a vendor-reported datapoint, not a tested result.

What a page actually costs

LlamaParse runs on credits. The published rate is 1,000 credits for $1.25, or $0.00125 per credit. Plans start at free: $0 a month for 10,000 credits, which LlamaIndex says covers about 1,000 pages. Starter is $50 a month for 40,000 credits, with pay-as-you-go up to 400,000 credits. Pro is $500 a month for 400,000 credits, plus a limited-time one-off bonus of 800,000 credits the company values at $1,000. Enterprise is quoted, not listed.

The per-page cost depends on the tier, because the current v2 API charges an extract tier plus a parse tier per page:

OCRia-1 powers the cost-effective and agentic tiers on the current "latest" version, and LlamaIndex says it now powers all LlamaParse modes at the same price — the upgrade came with no rate increase. There is also an Auto Mode that routes each page to a tier automatically, which LlamaIndex says can save up to 80% against running everything at the top tier.

Now the fine print. LlamaIndex's homepage says OCRia-1 replaces custom parsing pipelines "for less than half a cent per page." That figure traces to the older v1 rate card, where the published price was $0.003 a page. On the current v2 credit table, the cheapest full parse works out to a cent a page, and the agentic tier most teams will pick for complex documents works out to just over three cents. A cent a page is still cheap — 10,000 pages for $100 on the cost-effective tier — but it is double the advertised number, and anyone budgeting a million-page archive should use the credit table, not the headline.

Two more costs hide in the small print. Retained files — documents stored without an expiration — are billed at 100 credits per gigabyte per day, so a large archive left sitting in LlamaCloud keeps charging. And the free plan's 10,000 credits a month sounds generous until you run the agentic tier: at 25 credits a page, the free allowance covers 400 pages, not 1,000. The "about 1,000 pages" figure assumes the cheapest tier.

How OCRia-1 compares

The table makes the trade-off plain. Mistral OCR is less than half the price of LlamaParse's cheapest tier per page, and about an eighth of the agentic tier, but it trails by 20 points on ParseBench and by far more on charts. The hyperscalers are cheaper still for plain text detection, then charge separately — and steeply — for tables and forms, the exact features where OCRia-1 is strongest. Open-source options cost nothing per page and everything in engineering time: you host them, tune them and live with their errors. OCRia-1's case is that it is the only option in the table that leads on tables and charts while charging a flat per-page credit rate with no separate table surcharge. Whether that is worth the premium over Mistral depends on how much of your workload is tables and charts rather than plain text. It is also worth noting that parsing is only one half of the data problem: Google's Data Agent Kit takes a different approach, packaging MCP tools that let agents query structured data services directly instead of parsing documents at all.

LlamaIndex's own numbers put its scale at more than 1 billion documents processed and 300,000-plus LlamaParse users — company-reported figures, but they explain why this launch lands differently than one from a startup with no installed base: OCRia-1 shipped straight into an existing service with existing users, at the same price, rather than as a product that still has to find its market.

What OCRia-1 cannot do, yet

The limits are worth stating plainly, because the launch post does not dwell on them. First, there are no downloadable weights and no license. Launch coverage notes that the announcement mentions no open weights, and the practical consequence is that OCRia-1 is a hosted upgrade, not a model you can run. LlamaIndex says you can "deploy it in your environment," but that means through LlamaParse's enterprise arrangements, not a download. Teams that need air-gapped parsing or a model they can fine-tune will have to look elsewhere, or wait for a weights release that has not been promised.

Second, the benchmark story has a hole. All of OCRia-1's headline numbers come from LlamaIndex's own post, and no independent evaluation has been published yet. The olmOCR-bench figures soften the story rather than confirming it: a 1.2-point overall lead over olmOCR-2, not the 12-point lead ParseBench suggests. Until a third party runs the model on its own test set, the 87.4 should be treated as the company's best case, not the field's consensus.

Third, OCRia-1 is a parser, not a decision-maker. It extracts and structures what a document says; it does not judge whether a contract clause is risky, whether a financial statement is fraudulent, or whether an invoice is legitimate. Those judgments belong to the agent or the human downstream, and any pipeline that treats a confident parse as a verified fact is building on sand. The agents that consume parsed documents need their own safeguards too — Cognition's free Devin plugin, which patches the vulnerabilities agents introduce, is one example of the tooling that has to sit around a parser. The model also inherits the usual limits of document AI: handwriting is supported, but messy handwriting on low-quality scans is exactly the kind of input where every parser's accuracy drops, and LlamaIndex publishes no separate accuracy figure for handwriting.

Finally, the price story cuts both ways. At a cent a page on the cheapest tier, OCRia-1 is affordable for most workloads, but the agentic tier at 3.1 cents a page is eight times Mistral OCR's standard rate. Teams parsing millions of pages of plain text will find the cheaper options hard to beat; teams parsing thousands of pages of dense financial tables will find the premium easy to justify. The right tier depends on the documents, and the Auto Mode routing is worth testing before committing to one.

Pros and cons

Pros: Highest ParseBench score among the models LlamaIndex tested, with a 12.1-point lead; strongest table and chart parsing in the comparison set, including on three independent table benchmarks; converts charts into structured data rather than summaries; preserves formatting, math and checkboxes; no price increase — it powers LlamaParse at the same rates as before; generous free tier of 10,000 credits a month.

Cons: No downloadable weights or public license — hosted only; headline benchmark is the vendor's own, and the independent benchmark lead shrinks to 1.2 points; advertised "under half a cent a page" does not match the current credit table, where the cheapest tier is a cent a page; agentic tier costs 3.1 cents a page, eight times Mistral OCR's rate; no published latency or throughput guarantee; no separate accuracy figure for handwriting.

Frequently asked questions

What is OCRia-1?

OCRia-1 is LlamaIndex's 4-billion-parameter vision-language model for document parsing. It turns PDFs, scans, spreadsheets and slide decks into structured, LLM-ready output — text with layout preserved, tables rebuilt as tables, charts converted into data points, and math transcribed into LaTeX. It powers LlamaParse, LlamaIndex's commercial parsing service.

How much does OCRia-1 cost per page?

There is no separate OCRia-1 price; you pay LlamaParse's credit rates. Credits cost $1.25 per 1,000, and a page costs 8 credits on the cost-effective tier ($0.01), 25 on the agentic tier ($0.03125), 60 on Agentic Plus ($0.075) or 35 on Turbo ($0.04375). The free plan includes 10,000 credits a month. LlamaIndex advertises under half a cent a page, a figure that matches the older v1 rate card, not the current v2 credit table.

Is OCRia-1 open source? Can I download the weights?

No. The launch announcement mentions no open weights and no license, and launch coverage describes it as a hosted upgrade rather than a downloadable model. You use it through LlamaParse, hosted or deployed in your environment through LlamaIndex's enterprise arrangements. If you need a model you can run yourself, olmOCR-2 and Tesseract are open options.

How does OCRia-1 compare to Mistral OCR?

On ParseBench, OCRia-1 scores 87.4 against 67.5 for Mistral OCR 4.1, with the widest gap on charts. On price, Mistral is cheaper: $4 per 1,000 pages ($0.004 a page, $0.002 in batch) against $10 per 1,000 pages on LlamaParse's cheapest tier. The choice is accuracy on tables and charts against per-page cost.

Does OCRia-1 handle handwriting?

Yes, LlamaParse lists handwritten text among the document elements OCRia-1 parses, alongside tables, charts and complex layouts. LlamaIndex publishes no separate accuracy figure for handwriting, so messy handwriting on poor scans should be tested on your own documents before you commit a pipeline to it.

Key takeaways

• OCRia-1 is a 4-billion-parameter document-parsing model from LlamaIndex, introduced October 22, 2026, and it now powers LlamaParse at the same price as before.
• It scores 87.4 on LlamaIndex's own ParseBench, 12.1 points ahead of the next model, but the lead on the independent olmOCR-bench is only 1.2 points — read the headline number as a vendor claim, not a verdict.
• Tables and charts are its real edge: 91.5 on ParseBench tables, 85.1 on charts, with a 19-point lead over the next model on chart data extraction.
• A page costs $0.01 on the cost-effective tier, $0.03125 on agentic and $0.075 on Agentic Plus — double the advertised "under half a cent" on the cheapest tier.
• There are no downloadable weights and no public license, so OCRia-1 is a hosted upgrade, not a model you can run yourself.

Document parsing is the unglamorous step every AI agent pipeline depends on, and OCRia-1 is a genuine step forward on the hardest parts of it — tables and charts — from a vendor with the installed base to make it matter immediately. The benchmark lead is real but smaller than the headline suggests, the price is higher than the advertising, and the model is not yours to run. As of October 26, 2026, that is the honest state of OCRia-1: the strongest parser in LlamaIndex's own test, at a fair price, behind a hosted wall.