Qwen OCR vs Mistral OCR Pricing

Short answer: Alibaba's qwen-vl-ocr costs about $0.16 per 1,000 US Letter invoice pages on its Global rate card. Mistral OCR 4.1 costs $4.00 per 1,000 pages, or $2.00 through the Batch API. That makes Mistral roughly 25 times dearer at list and about 13 times dearer batched. Mistral earns some of that back by accepting PDFs directly, which the Qwen OCR model will not do at all.

These two get compared constantly, and almost every comparison you will find quotes one of them at the wrong price, usually because it is reading a regional rate card that does not apply or a model card for a different product. Here is what each one actually bills a US buyer, checked against the vendors' own pricing pages in August 2026.

The two products are not priced the same way at all

Mistral bills per page. One invoice page is one page, whether it arrived as a 150 DPI phone photo or a 300 DPI flatbed scan, and the rate never moves. Alibaba bills per token, and an invoice page becomes tokens according to how many pixels you send. That single difference drives almost everything else in this comparison.

Qwen's image tokenizer works on 32 by 32 pixel blocks. A US Letter page at 150 DPI is 1,275 by 1,650 pixels, which rounds up to 1,280 by 1,664 and tokenizes to 2,080 visual tokens. The same page at 300 DPI is 2,550 by 3,300 pixels and tokenizes to 8,058 tokens, nearly four times as many. On Mistral that difference costs you nothing. On Qwen it is most of your bill.

The Qwen figures assume roughly 40 prompt tokens in and about 900 tokens out per invoice, which is a realistic return for header fields plus a dozen line items. The arithmetic at 150 DPI: 2,120 input tokens times 1,000 pages is 2.12M input tokens at $0.043 per 1M, which is $0.09, plus 0.9M output tokens at $0.072 per 1M, which is $0.06. Total about $0.16.

Alibaba publishes four regional rate cards and the SDK samples point at the expensive one

This is the trap. Alibaba Model Studio prices qwen-vl-ocr at $0.043 input and $0.072 output per 1M tokens under Global scope, which is what the US (Virginia), Germany (Frankfurt) and China (Beijing) endpoints bill. The Singapore endpoint, tagged "International," charges $0.07 input and $0.16 output. That is 63% more on input and 122% more on output for identical work.

Alibaba's own SDK samples default to ap-southeast-1, which is Singapore. A team that copies the quickstart and never revisits it pays about $0.29 per 1,000 pages instead of $0.16, and nothing in the response tells them. If you are running Qwen for document work from the United States, check which region your client is pointed at before you benchmark anything.

Does Mistral OCR support PDF?

Yes. Mistral's OCR endpoint takes PDFs directly and returns text with layout structure, and OCR 4 added paragraph-level bounding boxes and structural block labels. Qwen's qwen-vl-ocr does not accept PDF at any size. It takes BMP, JPEG, PNG, TIFF, WEBP and HEIC below 4K resolution, and only JPEG and PNG between 4K and 8K, with a 10 MB ceiling.

For AP work that gap is larger than it looks. Vendor invoices arrive as PDFs, frequently multi-page ones. Using Qwen means running a rasterizer in front of the API, deciding a DPI for every document, splitting multi-page files into individual images, and paying for the compute that does it. That pipeline is not hard, but it is real engineering that Mistral's price includes and Qwen's does not. When you are modeling this, count the rasterizing instances and storage alongside the token spend, because per-unit API rates have a way of hiding inside a cloud bill that nobody is tracking against the actual document volume until the quarter closes.

How much does Mistral OCR cost with structured output?

Mistral runs three separate document meters, and picking the wrong one is an easy way to overpay or underdeliver. Plain OCR through mistral-ocr-latest is $4.00 per 1,000 pages. Document AI, which returns structured annotations against a schema you define, is $5.00 per 1,000 pages. The Libraries tool OCR meter is $3.00 per 1,000 pages. The Batch API halves whichever one you are on.

For invoice extraction the meter you almost certainly want is Document AI at $5.00, because plain OCR returns text and you still have to work out which string is the invoice number and which numbers belong to which line. That is the same distinction we walk through in invoice OCR versus AI extraction, and it is where most homegrown pipelines stall.

Can Qwen OCR return invoice fields as structured data?

Not for US documents. qwen-vl-ocr ships a built-in structured extraction task, but its trained document types are Chinese and Macau passports, Mainland Travel Permits, Home Return Permits, Chinese driver's licenses, vehicle nameplates, certificates of conformity, vehicle registration certificates and motor vehicle sales invoices. There is no US invoice type, no W-9 and no 1099. You can still prompt for fields as free text, but you are parsing model output yourself rather than calling a typed endpoint.

Two more constraints matter at volume. The system message on qwen-vl-ocr is fixed and cannot be customized, and default max output is 4,096 tokens. Raising it to 8,192 requires emailing an Alibaba commercial manager with your account ID, image type, QPS and daily volume. A dense multi-page invoice with forty line items can run into that ceiling.

Is Mistral OCR open source?

No. Mistral open-weights several of its language models, which is why this question comes up so often, but the OCR model is not among them. It is available only as a hosted API. Qwen is the opposite in an interesting way: the Qwen3-VL family is genuinely open under Apache 2.0, from 2B up to 235B parameters, but qwen-vl-ocr itself is a hosted service built on that architecture and is not published as weights. So neither vendor lets you self-host the specific OCR model you would be buying.

If self-hosting is the actual goal, the honest comparison is Qwen3-VL open weights against other open document models, not against either hosted OCR endpoint. The 30B-A3B checkpoint is the practical pick for document work, since it activates about 3B parameters per token and fits on a single H100.

Which is more accurate on invoices?

Neither vendor publishes field-level accuracy on invoices, and no independent benchmark we can find tests both on a US invoice set. Anyone quoting you a precise accuracy percentage for this matchup is quoting something they did not measure. What does exist is Alibaba's own caveat in the qwen-vl-ocr documentation, which states the model "may hallucinate if text in an image is too small or has a low resolution."

Read that next to the pricing and the tension is obvious. The cheapest way to run Qwen is to send 150 DPI images, and the vendor's own warning says low resolution is when the model invents text. On a $0.16 per 1,000 pages budget you can afford 200 DPI at $0.22 and still be an order of magnitude under Mistral, so there is little reason to run at the floor.

Which one should you actually buy?

Pick Qwen if you are processing high volume, you already have image preprocessing in your stack, you can live without typed invoice output, and someone owns the pipeline. At a million pages a year the gap between $0.16 and $4.00 per 1,000 is roughly $3,800, which is real but is also about three weeks of one engineer's time. Below a few hundred thousand pages the savings rarely cover the build.

Pick Mistral if you want to send a PDF and get structured fields back without building anything, or if your documents are multi-page, or if you need annotations against your own schema. The $5.00 Document AI rate buys a pipeline you would otherwise write and maintain yourself.

Pick neither if your job is specifically accounts payable rather than general document parsing. Both of these are text and layout engines that hand you output to map; neither knows what a purchase order number is or which total is subject to tax. A purpose-built invoice OCR API returns typed header fields and line items at a flat monthly rate, with no rasterizing, no region selection and no token math. Full rate cards for both vendors, including every meter and regional tier, sit on our Qwen OCR pricing and Mistral OCR pricing pages, and Mistral OCR against Azure Document Intelligence covers the enterprise alternative.

Rates verified against mistral.ai/pricing/api and the Alibaba Cloud Model Studio pricing documentation in August 2026. Vendor pricing changes; check the source before you commit to a contract.