Document parsers for RAG: what they do with tables and scans and what a page costs (October 2026)

A thousand pages cost $1.25 to $56.25 at one vendor, and $0 in license at three

The widest spread sits inside a single product. LlamaParse bills credits at $1.25 per 1,000 credits, and one page costs 1 credit on the Fast tier, 3 on Cost-effective, 10 on Agentic and 45 on Agentic Plus. That is $1.25, $3.75, $12.50 and $56.25 per 1,000 pages. Layout extraction adds 3 credits to any tier, another $3.75.

Amazon Textract prices by feature, per 1,000 pages in the Oregon region: $1.50 for plain text and handwriting, $15 for Tables, $50 for Forms, $15 for Queries and $4 for Layout. Datalab, the company behind Marker, lists $4 for its hosted fast or balanced conversion and $10 for accurate. I took the Textract numbers from AWS's published price list, dated September 11, 2026, because the pricing page loads its prices by script. I left Mindee out of the table because its documentation and pricing page disagree on the currency (details at the end).

The table shows each price next to the license and whether you can run the tool yourself. LlamaParse and Textract are hosted only. The calculator multiplies your monthly pages by the same list prices. At 20,000 pages a month, LlamaParse costs $25 on Fast and $1,125 on Agentic Plus, and Textract costs $30 for text and $300 with Tables. Plan fees and free allowances are not included.

What this means for you: compare tools at the same job, not at the headline price. A $1.50 plain-text page and a $15 page with tables are different products, and the 45 times gap inside LlamaParse shows that much of the bill is the effort level you ask for, not the vendor.

Sources: [1][2][3][4][5][6]

Free allowances differ more than prices do

LlamaParse's Free plan includes 10,000 credits. That buys 10,000 Fast pages, 3,333 Cost-effective pages, 1,000 Agentic pages or 222 Agentic Plus pages. Starter is $50 a month for 40,000 credits and Pro $500 for 400,000. Parsed files are cached for 48 hours, and re-parsing the same file in that window is free.

Textract's free tier lasts three months for new AWS customers. It covers 1,000 pages a month of text extraction but only 100 pages a month with Forms, Tables and Layout. Datalab's Free plan credits $20 a month on a work email and $10 on a personal one, which is 5,000 pages of fast conversion at $4 per 1,000 pages or 2,000 pages at the accurate rate. Mindee offers a 14-day trial with 200 pages.

Mindee bills credits, 1 per page by default. Starter costs €44 a month with annual billing for 6,000 credits a year, about €0.09 per page if you use every credit, and extra credits cost €0.044 each, or €44 per 1,000 pages. The plans page says a larger credit bundle lowers the credit price, so the overage rate is the number to hold against Textract's $50 for Forms. Mindee's price is in euros, the others' in dollars, and I did not convert.

What this means for you: a free tier tests quality on a sample, not on a corpus. A new AWS account covers about 100 table pages a month, and testing 500 table pages on Textract then costs about $6 for the remaining 400 (400 pages at $0.015).

Sources: [7][3][2][4][8][9][10]

Tables mean three different things in the documentation

Docling lists table structure next to page layout and reading order and exports to Markdown, HTML and JSON. Marker says it rebuilds tables from the PDF text layer, takes scanned tables from full-page OCR and falls back to its vision model when its confidence is low.

Mindee works the other way round. Its documentation says a repeated table, such as the line items of an invoice, is extracted as a list of nested objects, which returns structured fields rather than a table-layout export. That suits an invoice pipeline. It is a different output from a Markdown table in a RAG index.

Amazon Textract sells tables as a priced feature. Its FAQ says tables work best when they are visually separated from the rest of the page, not laid over an image or pattern, and when the text in them is upright. LlamaParse's Fast tier returns spatial text only and no Markdown, according to its pricing documentation, so a table there is not a Markdown table. Marker's optional LLM mode merges tables that run across pages.

What this means for you: before you pay for a tier, run ten of your own table pages through it and check whether rows and columns survive. A price per page says nothing about that.

Sources: [11][12][13][14][3]

Scans: the only published score for old scans is 43.2, and it comes from a competitor

Datalab's repository is the only one of the six that publishes per-category scores. In its own run on olmocr-bench, a test set of 1,403 PDFs with about 8,400 pass/fail checks, Marker's balanced mode scores 73.4 on tables, 99.7 on the baseline category and 43.2 on old scans. Its overall score is 76.0, against 72.7 for MinerU's pipeline backend and 50.3 for Docling's default pipeline.

The modes explain part of that. Marker's balanced mode re-reads a whole page with its vision model whenever any of the embedded text is bad. Its fast mode repairs single garbled blocks and makes a full-page pass only for pages that are scanned or mostly bad. For a PDF whose text layer is wrong, the README suggests forcing OCR on all pages.

This is a vendor benchmarking two other tools, so I treat it as a claim and not as a ranking. It still shows where the weak spot is. Scans are the weakest category even in the best mode, and with OCR switched off Marker scores 14.3 on old scans. I did not check whether Docling or MinerU publish scores on the same set.

The hosted services state limits instead of scores. Amazon Textract asks for scans at ideally 150 DPI or more, reads handwriting only in English and reads printed text in English, German, French, Spanish, Italian and Portuguese. Mindee's documentation says scanned PDFs generally work better than phone photos. Docling lists six OCR engines: RapidOCR, Nemotron-OCR, EasyOCR, ocrmac, tesseract-CLI and tesserocr. MinerU says it parses scanned and captured document images.

What this means for you: if your archive is old scans or phone photos, plan for the expensive tier or a GPU, and judge the tool on your 20 worst pages, not on an average one.

Sources: [12][14][13][15][16]

Self-hosting: the license is free, the machine is not

Docling is MIT licensed and hosted by the LF AI & Data Foundation. Its README lists local execution for sensitive data and air-gapped environments, and its documentation shows how to download the layout and table models in advance for offline use. Version 2.133.0 was released on October 3, 2026.

MinerU's license is Apache 2.0 with added conditions. A separate commercial license is required once a product passes 100 million monthly active users or $20 million in monthly revenue, and anyone offering an online service built on it must say so prominently. At the basic tier it runs on CPU with 2 GB of RAM. The standard and advanced tiers on a GPU ask for 16 GB of RAM and 8 GB or more of VRAM. It parses locally by default and sends a document out only when you pass the remote flag. Its online service at mineru.net publishes no prices as of October 5, 2026.

Marker's code is Apache 2.0, but its model weights use a modified OpenRAIL-M license that is free for research, personal use and startups under $5 million in funding or revenue. Above that, Datalab's commercial terms apply. On one B200 GPU Datalab measures 2.9 pages per second in balanced mode, about 10,400 pages an hour, against 0.54 for MinerU and 2.1 for Docling in the same run. Datalab publishes no GPU cost, so I did not turn this into a price.

GitHub showed on October 5, 2026 about 81,100 stars for MinerU, 68,400 for Docling and 40,200 for Marker, with the latest code changes on September 30 for MinerU and October 2 for Marker. That tells you about attention and upkeep, not about quality on your files.

What this means for you: under those size limits and with a GPU, all three cost the price of the machine. A larger company should read the license first. Docling's MIT license has no such threshold, while for MinerU and Marker the threshold is a contract question.

Sources: [5][17][6][16][18][12]

Which one fits which job, and for whom it does not

I sorted the six by three questions: can the files leave your network, do you need fixed fields or the full text of a page, and do you already run on AWS. The helper below walks through them and ends on a starting point.

LlamaParse gives a ladder of tiers and an EU region endpoint, so it suits a team that wants one hosted API and can pick the tier per document type. Firecrawl and Jina Reader also read PDFs. Firecrawl bills 1 credit per PDF page on top of 1 credit per scrape, so a 10-page PDF costs 11 credits, and Jina Reader says it parses PDF URLs with PDF.js. Both start from a URL, and I left them out of the comparison for that reason.

Sources: [3][19][20][21][14][10]

How I counted

On October 5, 2026 I read the pricing pages, documentation and repository READMEs of Docling, MinerU, LlamaParse, Marker, Mindee and Amazon Textract, plus AWS's price list file. I counted 6 of the 9 tools in the niche, compared license, hosting and the statements on tables and scans, and pulled every price from the vendor itself. How we review explains the general method. I did not upload a file or create an account, so none of this measures accuracy on your documents.

Three sources disagreed. Mindee's pricing page shows $44 and $116 a month for Starter and Pro, while its documentation plan page shows the same numbers in euros. The documentation also states that Starter includes 6,000 credits per year, not per month, and that extra credits cost €0.044 each, while the pricing page speaks of a 20% overage surcharge. I took the documentation figures and corrected our entry. Datalab's README promises $5 in free credits, its pricing page $20 a month on a work email. I took neither. My first read of the Textract page also listed Layout as free, but the price list shows $0.004 a page, so I took the price list.

I left out Crawl4AI, which crawls websites, and one price I could not find: MinerU's online service publishes none.

Sources: [22][8][9][4][1][23]