DeepSeek OCR Pricing: API Cost, Self-Hosting Cost and DeepSeek OCR 2 Rates Per 1,000 Pages

The three ways to pay for DeepSeek OCR, side by side

Almost every "DeepSeek OCR pricing" search is really one of three different questions, and they have three different answers. The table below separates them. Rates were read on 21 August 2026 from DeepSeek's own API documentation and from Lambda's published on-demand GPU list.

The distinction in row two is the single most misunderstood thing about DeepSeek OCR pricing, and it is worth being blunt about it: there is no DeepSeek OCR endpoint on the DeepSeek API. The models list at api-docs.deepseek.com carries exactly three entries, deepseek-v4-flash, deepseek-v4-pro and deepseek-v4-flash-vision-exp. None of them is DeepSeek-OCR. If you want the OCR model, you run it.

What self-hosting DeepSeek OCR actually costs per 1,000 pages

DeepSeek published its own throughput figure in the DeepSeek-OCR paper: the model "can generate training data for LLMs/VLMs at a scale of 200k+ pages per day (a single A100-40G)". That is 8,333 pages an hour on one GPU. Lambda lists the A100 40 GB at $1.99 an hour on demand. Divide one by the other and you get the marginal cost of a page.

The number that matters, though, is not the marginal cost. It is what you get billed, because GPUs rent by the hour and a small AP department cannot keep one busy. Here is both.

Read the bottom row carefully, because it is the honest ceiling: one rented A100 running around the clock for a full month costs about $1,453 and can chew through roughly six million pages. Almost nobody in accounts payable has six million pages. A US mid-market company running 10,000 vendor invoices a month at two pages each needs the GPU for about two and a half hours.

DeepSeek OCR pricing on the API: the model you can buy is not the model you want

If you go looking for a hosted DeepSeek OCR API price, you land on deepseek-v4-flash-vision-exp, because it is the only DeepSeek model that accepts images. It is genuinely cheap. It is also, for invoices specifically, the wrong tool, and the documentation says why in a sentence most people skim past.

Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC. Everything else is off-peak at half price. There is a quiet win for American buyers in that schedule: converted to Eastern time those peak windows fall roughly between 9pm and midnight and again between 2am and 6am, so a US team running invoices during its own working day is off-peak all day. A team in Singapore or Frankfurt is not.

Now the catch. DeepSeek's vision guide states that before inference "larger images are scaled down while preserving their aspect ratio, so that the total pixel count after resizing is roughly that of an 800x800 image," with "an upper bound of 384 tokens per image." Run that resize against a real scanned invoice:

The first two rows are the finding. A 300 DPI scan and a 200 DPI scan of the same US Letter invoice arrive at the model as the identical 703 by 910 image, so paying for better scanning hardware buys you nothing on this endpoint. Roughly 83 DPI is well under the 300 DPI that document capture vendors have recommended for decades, and it is the resolution at which six point remittance terms, hyphenated part numbers and the difference between an 8 and a 3 in a line-item total start to go. The docs also list the accepted formats as JPEG, PNG, GIF and WebP. PDF is not on that list, so multi-page vendor bills have to be rasterized page by page before they can be sent at all.

None of this makes the vision model bad. It is a general-purpose model priced like one, and for reading a screenshot or a chart it is excellent value. It is simply not a document capture engine, and its price should not be quoted as DeepSeek OCR pricing.

DeepSeek OCR vs DeepSeek OCR 2: what changed, and the license change nobody mentions

DeepSeek released DeepSeek-OCR on 20 October 2025 and DeepSeek-OCR 2 on 27 January 2026. Both are free to download. The differences that affect a buying decision are below, taken from the two repositories, the two papers and the Hugging Face model metadata.

The license line deserves a moment in a US procurement review. Version 1 is MIT, which is about as permissive as software licensing gets. Version 2 is Apache 2.0, which is still permissive but carries an express patent grant and its own attribution and NOTICE obligations. If your legal team maintains an approved-license list, or if you redistribute the model inside a product you ship to customers, upgrading from version 1 to version 2 is not a silent change and it is worth a five minute conversation before it lands in your build.

The second thing worth noticing is that DeepSeek-OCR 2 dropped the fixed resolution ladder. Version 1 let you buy accuracy by the token: 64 tokens on a clean typed page, 400 on a dense one. Version 2 decides for you within a 256 to 1,120 range. For a research workload that is a simplification. For a cost model built on tokens per page it removes a lever you may have been counting on.

The compression numbers, and what they mean for a line-item total

The DeepSeek-OCR paper is unusually candid about the accuracy trade. Its own words: "when the number of text tokens is within 10 times that of vision tokens (i.e., a compression ratio < 10x), the model can achieve decoding (OCR) precision of 97%. Even at a compression ratio of 20x, the OCR accuracy still remains at about 60%."

97% character precision is a strong result for an open 3.3 billion parameter model and it is not the same thing as 97% of invoices coming out correct. An invoice has perhaps forty values on it that a bookkeeper cares about, and a total that is wrong by one digit is not 97% right, it is wrong. This is the reason serious AP workflows put a validation layer over any OCR engine: totals that reconcile to line items, dates that parse, vendor names that match a master list, tax that recomputes. That layer is where the engineering budget goes, and it is identical whether the characters came out of DeepSeek, Tesseract or a commercial API.

The paper's benchmark claims are about efficiency rather than raw accuracy, and they are impressive on that axis: on OmniDocBench, DeepSeek-OCR "surpasses GOT-OCR2.0 (256 tokens/page) using only 100 vision tokens, and outperforms MinerU2.0 (6000+ tokens per page on average) while utilizing fewer than 800 vision tokens." Fewer tokens for the same or better output is exactly what drives that $0.24 per 1,000 pages.

What hardware you need, and whether it runs without a GPU

The weight files answer most of this. Both models are single-file BF16 checkpoints, so the arithmetic is 2 bytes per parameter.

Those VRAM figures are weights plus room for activations and a batch, not a vendor specification. The repositories document a CUDA 11.8 and PyTorch 2.6.0 environment and install flash-attn 2.7.3, which is CUDA only. Both models are supported by vLLM, and DeepSeek-OCR was merged into upstream vLLM on 23 October 2025, which is what makes the batch throughput number reachable in practice rather than in theory.

Here is the cost comparison that matters if you are choosing between renting and buying, using Lambda's published on-demand rates read on 21 August 2026:

For this specific job the A100 40 GB is the right rental, which is convenient, because it is also the card the published throughput figure was measured on, so the cost model does not require you to guess at a scaling factor.

How DeepSeek OCR cost compares to the commercial invoice APIs

This is the comparison people are really running, so here it is in one place. Commercial rates below are the figures published by each vendor and re-verified by us in August 2026.

Column three is the whole argument. DeepSeek-OCR is roughly forty times cheaper per page than Textract and it does a different job. It converts a page image into markdown. It does not hand you an invoice number, a vendor, a due date, a tax total and a list of line items with quantities and unit prices, keyed and typed, ready to post. Turning good markdown into good structured invoice data is a second system: prompt or parse the markdown into fields, normalize dates and currencies, reconcile line items to the total, match to a purchase order, decide what a human has to look at. That system is where the real money goes, and its cost does not fall just because the OCR got cheap.

So the honest way to read the table is not "DeepSeek wins on price." It is that the OCR layer has become close to free and the value moved up the stack. If you have GPU infrastructure and engineers, DeepSeek-OCR is an outstanding component and you should use it. If what you need is invoice data in a spreadsheet or your accounting system on Monday, the per-page rate is not the number you are optimizing.

When self-hosting DeepSeek OCR is genuinely the right call

Three situations make the build worth it, and it is worth being specific rather than encouraging.

Volume above roughly 100,000 pages a month. At that point the commercial per-page rate is a real line item. 100,000 pages at $10 per 1,000 is $1,000 a month; the same volume on a rented A100 is about $24 of GPU. If you already employ the engineers, the payback on building is measured in months.

Data that cannot leave your building. Open weights on your own hardware is a genuinely different compliance posture from any API, and for defense contractors, healthcare adjacent finance teams and anyone under a contractual data residency clause it can be the only acceptable answer. Note that this is an argument for self-hosting, not for DeepSeek specifically, and that hosting a Chinese-developed model may itself raise questions in some US procurement processes even though the weights are permissively licensed and run entirely offline.

Documents nobody sells a parser for. Prebuilt invoice models are trained on invoices. If your documents are field tickets, bills of lading with handwritten annotations, or a format your industry invented in 1987, a general document model you control may beat a commercial parser that was never trained on your paper.

Outside those three, the arithmetic usually goes the other way. A three person AP team processing 800 invoices a month saves about $8 of GPU versus a commercial API and takes on a model server, a queue, a rasterization step, a parsing layer, a validation layer and an on-call rotation. That is not a saving, it is a hobby with a budget code.

Last updated

Last updated August 2026. Every figure on this page was read from a primary source on 21 August 2026: DeepSeek's models and pricing page and vision guide at api-docs.deepseek.com, the DeepSeek-OCR and DeepSeek-OCR-2 repositories on GitHub, the two arXiv papers, the Hugging Face model metadata for parameter counts and licenses, Mistral's API pricing page, and Lambda's published on-demand GPU rates. Model prices change without notice, and DeepSeek's own documentation says so explicitly. Check the source before you commit a budget to any of it.