Qwen OCR vs Azure Document Intelligence Pricing

Short answer: Alibaba's qwen-vl-ocr costs about $0.16 per 1,000 US Letter invoice pages and Azure Document Intelligence's prebuilt invoice model costs $10.00 per 1,000. Qwen is roughly 60 times cheaper and returns raw text. Azure is 60 times dearer and returns typed invoice fields with line items at 87% published accuracy. You are not choosing between two prices, you are choosing whether to build the mapping layer yourself.

Both vendors publish real numbers, which already puts them ahead of most of this market. The difficulty is that they publish them in different units. Alibaba bills tokens, Azure bills pages, and until you convert one into the other the comparison is meaningless. Here is that conversion, plus the four things about each service that only bite buyers in the United States.

The one-line comparison

How a token rate becomes a page rate

Qwen bills images as visual tokens, and one visual token is a 32 by 32 pixel block. Page dimensions round up to a multiple of 32, so a US Letter page at 150 DPI is 1,275 by 1,650 pixels, tokenized as 1,280 by 1,664, which is 2,080 visual tokens. Add roughly 40 tokens of prompt and assume about 900 output tokens for a full transcription, and 1,000 pages works out to 2.12 million input tokens and 900,000 output tokens. At the Global-scope rate that is nine cents of input and six cents of output.

Resolution is therefore a billing lever, and unlike some competing vision endpoints it is a lever that actually works. Qwen's default ceiling is 8,388,608 pixels, or 8,192 tokens, and a 300 DPI Letter page comes in at 8,415,000 pixels, only 0.3% over. Your scan arrives essentially intact rather than being crushed to something unreadable.

Azure does not care about resolution. A page is a page, which makes budgeting trivial and makes high-resolution archival scans free to process. Going from 150 to 300 DPI multiplies the Qwen bill by about 2.6 times and changes the Azure bill by nothing at all.

The region trap, and why it only affects one of them

Alibaba publishes four separate rate cards: Singapore, China (Beijing), Germany (Frankfurt) and US (Virginia). The SDK quickstart points at Singapore, and Singapore is the most expensive of them for OCR work. On qwen-vl-ocr the Singapore rate is $0.07 in and $0.16 out, against $0.043 and $0.072 on the Global rate card. Same model, same output, and nothing in the API response warns you that you picked the expensive endpoint.

There is a harder version of this problem. Model IDs carrying a -us suffix pin inference to the United States, and there is no qwen-vl-ocr-us. The dedicated OCR model exists only under Global scope, where compute is scheduled worldwide. If your compliance position requires that document content stays in the United States, qwen-vl-ocr is not available to you at any price. The newest model in the family, qwen3.5-ocr, is mainland China only, so the version Alibaba's own docs recommend migrating to cannot be bought by a US business at all.

Azure has the opposite shape. You choose a region, the data stays there, and if that is still not enough you can run the model in a connected container inside your own network, or in a fully disconnected air-gapped deployment. The pricing detail almost nobody publishes is that air-gapped costs exactly the same as connected once you annualize it: the disconnected prebuilt tier is $8,640 a year for 100,000 pages a month, and the connected tier is $720 a month, which is the same $8,640. Microsoft charges no premium for the air-gapped option.

What each one actually returns

This is the part the price comparison hides. Qwen OCR returns text. Good text, in sensible reading order, but it does not know that the number in the bottom right is the invoice total rather than the subtotal, the freight line or a prior balance. Everything between that text and an AP entry is code you write and maintain.

Azure's prebuilt invoice model returns named fields: vendor name and address, customer details, invoice ID, invoice date, due date, purchase order number, subtotal, tax, total, amount due, and a line items array with description, quantity, unit price and amount on each row. Published third-party testing on scanned documents put it at 93% on header fields and 87% on line items, which is the strongest line item figure in that test by a wide margin. For comparison, Google Document AI scored 40% and a GPT-4o pipeline scored 57% on the same line items.

There is one more limitation worth knowing before anyone builds on qwen-vl-ocr for accounts payable. Its built-in structured extraction task is trained on Chinese document types: mainland and Macau passports, travel permits, Chinese driver's licenses, vehicle registration certificates and motor vehicle sales invoices. There is no US invoice, W-9, 1099 or purchase order among them. The system message is fixed and cannot be customized, the model accepts no PDF input, files cap at 10 MB, and Alibaba's own documentation warns that the model "may hallucinate if text in an image is too small or has a low resolution." None of that makes it a bad OCR engine. It makes it an OCR engine rather than an invoice engine.

Where the volume breaks land

Qwen has no commitment tiers and no batch discount. The rate you pay at page one is the rate you pay at page one million, which is a fair deal when the rate is sixteen cents per thousand pages. There is a free quota of 1,000,000 tokens, but it applies only to Singapore deployment and expires 90 days after activation, so it is a trial rather than a tier.

Azure does have commitment tiers, and they are not on the pricing page. For the prebuilt invoice model they start at $190 a month for 20,000 pages, an effective $9.50 per 1,000, and a container commitment at the same volume is $152 a month, or $7.60. That is 24% below pay-as-you-go at a volume an ordinary mid-market AP department reaches. Overage bills at the tier's own effective rate rather than snapping back to list, so buying slightly under your real volume is cheap insurance.

What Azure does not give you is a batch discount. Every batch meter is priced identically to its synchronous equivalent, which is unusual in 2026 when Mistral, OpenAI and Anthropic all take 50% off batch work. If your AP volume is overnight and latency-insensitive, that is money Azure leaves on the table and you cannot recover.

So which one should you buy?

Buy Qwen OCR if you are building a pipeline, you already have or want to write the field-mapping and validation layer, your documents are images rather than PDFs, and no compliance rule forces your content to stay in the United States. At sixteen cents per thousand pages the reading step effectively disappears from your budget and the engineering time becomes the whole cost.

Buy Azure Document Intelligence if you need typed invoice fields and line items without building them, you need a named region or an air-gapped deployment, or your volume clears 20,000 pages a month and the commitment tiers apply. The $10 per 1,000 is not really an OCR price, it is the price of not writing the mapping layer, and for most AP teams that is the cheaper of the two options once a developer's time is in the model.

Buy neither if your problem is a monthly pile of vendor PDFs and a person retyping them. A cloud API is a component, not a workflow, and the pages get read either way. The thing worth watching in both cases is what the bill actually does over the first three months, because per-token pricing has a habit of surprising teams that budgeted from a spreadsheet, and pulling your cloud and SaaS spend into one read-only view catches that drift far earlier than a quarterly invoice review does.

How much does Qwen OCR cost compared to Azure Document Intelligence?

Qwen OCR costs about $0.16 per 1,000 US Letter invoice pages at 150 DPI on Alibaba's Global rate card, and Azure Document Intelligence's prebuilt invoice model costs $10.00 per 1,000 pages. Qwen is roughly 60 times cheaper, but returns unstructured text rather than the typed invoice fields and line items Azure produces.

Does Qwen OCR support PDF files?

No. qwen-vl-ocr accepts images only: BMP, JPEG, PNG, TIFF, WEBP and HEIC below 4K resolution, and JPEG or PNG only between 4K and 8K, with a 10 MB file cap. You have to rasterize PDF pages to images before sending them. Azure Document Intelligence accepts PDFs directly, which removes a rendering step and the resolution decisions that come with it.

Can I keep my documents in the United States with Qwen OCR?

Not on the dedicated OCR model. qwen-vl-ocr is published only under Global scope, where Alibaba schedules inference compute worldwide, and there is no qwen-vl-ocr-us variant. Only some Qwen3-VL model IDs carry a -us suffix that pins inference to the United States, and those cost significantly more per input token.

Is Azure Document Intelligence accurate enough for line items?

It is the strongest published figure available. Third-party testing on scanned documents in January 2025 measured Azure Document Intelligence at 93% on header fields and 87% on line items, against 40% for Google Document AI and 57% for a GPT-4o pipeline. Test on your own worst vendor layouts before committing, because line item accuracy varies far more by document than header accuracy does.

Getting from either one to an actual AP entry

Whichever engine you pick, the finished job is the same: typed fields, line items you can trust, and a file your accounting system accepts. If you would rather skip the pipeline entirely, our OCR accounts payable page lays out what the whole extraction step costs across every engine, and invoice line item extraction covers the part that decides whether a three-way match runs without a human. The full rate cards behind this comparison are on Qwen OCR pricing and Azure Document Intelligence pricing, both re-verified in August 2026.

Rates read from Alibaba Cloud Model Studio documentation and Microsoft's Azure Retail Prices API on 26 August 2026. Accuracy figures from published third-party testing by BusinesswareTech, January 2025.