Last updated November 1, 2026

Self-hosted OCR cost calculator

At 100,000 clean pages a month, run as a batch job, Chandra OCR (olmOCR-Bench 83.1) on a Vast.ai RTX 4090 costs $28.94 in GPU rental, against $150 for AWS Textract and $400 for Mistral OCR. The GPU price is a marketplace median. Its pages-per-hour rate is an estimate.

Enter your own page count below. The calculator picks an open model and a GPU, then compares that bill with Textract, Document AI, Azure, and Mistral.

By Tayyab Javed · Last updated November 1, 2026 · Prices checked November 1, 2026

$28.94

this month · $0.29 per 1,000 pages · or $295.24 if the GPU stays on

Chandra OCR

olmOCR-Bench 83.1

Vast.ai · NVIDIA GeForce RTX 4090
$0.4044/hr · 1,397 pages/hour, estimate · score 83.1 on olmOCR-Bench

Chandra OCR has the highest olmOCR-Bench overall (83.1) among models that cost under twice the cheapest GPU-hour bill.

Quality is a property of the model, not the GPU. The same model has the same score on an L4 and an H100. The GPU changes pages per hour and the bill. The pages-per-hour figure is an estimate, scaled from a published H100 measurement. The hourly price is the median of 60 Vast.ai offers, not a list price. Offers that day ran from $0.123 to $0.4711.

Published list price$37.13

PaddleOCR-VL
RunPod Secure Cloud A100 SXM 80GB

Measured speed$27.83

PaddleOCR-VL on H100 SXM on-demand median. Pages/hour was measured. The price is still a median.

Below the score floor$8.03

DeepSeek-OCR, score 75.7. Shown because it was measured. It is under the score floor for this document type, so it is not the pick.

Monthly cost at 100,000 pages

Self-host, this batch$28.94

Self-host, GPU left on$295.24

Textract$150.00

Document AI$150.00

Azure$150.00

Mistral OCR$400.00

Against the APIs

Textract$150.00Self-host costs less this month. Always-on breaks even at 196,829 pages.
Document AI$150.00Self-host costs less this month. Always-on breaks even at 196,829 pages.
Azure$150.00Self-host costs less this month. Always-on breaks even at 196,829 pages.
Mistral OCR$400.00Self-host costs less this month. Always-on breaks even at 73,811 pages.

You’d save $121.06 this month vs Textract. Want help setting it up?

Free, from this page

Get your OCR deployment plan (free)

One email with this page’s inputs, the recommended model, GPU, and provider, a deployment checklist, and whether self-host or the API is the smaller GPU bill.

Preview of the email

OCR deployment plan

100,000 pages · Clean PDFs

Chandra OCRVast.ai · NVIDIA GeForce RTX 4090

$28.94 this month

Textract $150.00 for the same pages. GPU rental only.

Pin the model commit and the GPU SKU, then time a few hundred of your own pages. The non-H100 rates here are bandwidth estimates.
Decide batch versus a warm GPU from the break-even, not from the hourly price alone.
Keep the PDFs in storage you control. A shared community GPU is the wrong place for a contract that forbids that.

Self-hosted OCR vs AWS Textract cost

At 100,000 pages a month, run as a batch job, Chandra OCR on a Vast.ai RTX 4090 is $28.94 of GPU rental, or $0.29 per 1,000 pages. Textract Detect Document Text is $150, or $1.50 per 1,000 pages. Kept on for all 730 hours, the same GPU is $295.24 a month, which equals the Textract bill at 196,829 pages.

On AWS itself, an always-on L4 (g6.xlarge, $0.8048 an hour in us-east-1, $587.50 a month) breaks even with Textract at 391,669 pages a month. Both numbers are GPU rental or API list price only. Engineering time, storage, egress, and retries are not in either one.

Best open-source OCR model: quality vs cost

Quality is a property of the model, not the GPU. The same model has the same score on an L4 and an H100. The GPU changes pages per hour and the bill. Default sort is the highest score for the selected document types, then the lower bill. Scores do not change when the GPU changes.

Each dot is that model’s cheapest setup for this workload. Managed APIs are not plotted: Textract, Document AI, and Azure are not on olmOCR-Bench, and the Mistral row on that bench is not labeled OCR 4.1.

SortModel

Every open OCR model and GPU setup for 100,000 pages a month: published score, hourly price, pages per hour, and monthly bill. Checked 2026-11-01.
Model	GPU	Provider	$/hour	Always-on / month	This workload	Per 1,000	Pages/hour	olmOCR-Bench	OmniDocBench	Category
Chandra OCR · recommended	NVIDIA GeForce RTX 4090	Vast.ai RTX 4090 24GB on-demand median	$0.4044 median	$295.24	$28.94	$0.29	1,397 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA A10 / A10G	Vast.ai A10 on-demand median	$0.2453 median	$179.09	$29.50	$0.29	832 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA A100 80GB SXM	Vast.ai A100 SXM4 80GB on-demand median	$1.1253 median	$821.49	$39.81	$0.40	2,827 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA H100 80GB SXM	Vast.ai H100 SXM on-demand median	$2.2044 median	$1,609	$47.47	$0.47	4,644 measured	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA A100 80GB SXM	RunPod Secure Cloud A100 SXM 80GB	$1.79 on-demand Community $1.39	$1,307	$63.33	$0.63	2,827 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA GeForce RTX 4090	RunPod Secure Cloud RTX 4090	$0.89 on-demand Community $0.34	$649.70	$63.69	$0.64	1,397 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA A100 80GB PCIe	RunPod Secure Cloud A100 PCIe 80GB	$1.79 on-demand Community $1.19	$1,307	$66.73	$0.67	2,682 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA L4	Vast.ai L4 on-demand median	$0.3378 median	$246.58	$81.22	$0.81	416 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA H100 80GB SXM	Modal Nvidia H100 SXM5	$3.9492 preemptible Non-preemptible $11.85	$2,883	$85.04	$0.85	4,644 measured	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA H100 80GB SXM	RunPod Secure Cloud H100 SXM	$3.99 on-demand Community $2.69	$2,913	$85.92	$0.86	4,644 measured	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA A100 80GB SXM	Modal Nvidia A100 80GB	$2.4984 preemptible Non-preemptible $7.4952	$1,824	$88.39	$0.88	2,827 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA H100 80GB PCIe	Vast.ai H100 PCIe on-demand median	$2.5422 median	$1,856	$91.69	$0.92	2,773 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA A100 40GB SXM	Lambda 1× A100 SXM 40GB	$1.99 on-demand	$1,453	$92.32	$0.92	2,156 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA H100 80GB SXM	Lambda 1× H100 SXM	$4.29 on-demand	$3,132	$92.38	$0.92	4,644 measured	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA GeForce RTX 4090	RunPod Serverless RTX 4090 worker	$1.34 on-demand	$978.20	$95.90	$0.96	1,397 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA A100 80GB SXM	RunPod Serverless A100 80GB worker	$2.72 on-demand	$1,986	$96.23	$0.96	2,827 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA H100 80GB PCIe	RunPod Secure Cloud H100 PCIe	$2.89 on-demand Community $1.99	$2,110	$104.24	$1.04	2,773 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA H100 80GB SXM	RunPod Serverless H100 80GB worker	$5.39 on-demand	$3,935	$116.06	$1.16	4,644 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA H100 80GB PCIe	Lambda 1× H100 PCIe	$3.29 on-demand	$2,402	$118.66	$1.19	2,773 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA A10 / A10G	AWS g5.xlarge	$1.006 on-demand	$734.38	$120.95	$1.21	832 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA A100 80GB SXM × 8	AWS p4de.24xlarge	$27.45 on-demand	$20,036	$121.38	$1.21	22,613 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA A10 / A10G	Modal Nvidia A10	$1.1016 preemptible Non-preemptible $3.3048	$804.17	$132.44	$1.32	832 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA L4	RunPod Secure Cloud L4	$0.59 on-demand Community $0.44	$430.70	$141.87	$1.42	416 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA H100 80GB SXM	AWS p5.4xlarge	$6.88 on-demand	$5,022	$148.15	$1.48	4,644 measured	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA A10 / A10G	Lambda 1× A10	$1.29 on-demand	$941.70	$155.09	$1.55	832 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA L4	Google Cloud g2-standard-4	$0.7068 on-demand	$515.99	$169.96	$1.70	416 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA A100 80GB SXM	Google Cloud a2-ultragpu-1g	$5.0688 on-demand	$3,700	$179.32	$1.79	2,827 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA L4	Modal Nvidia L4	$0.7992 preemptible Non-preemptible $2.3976	$583.42	$192.17	$1.92	416 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA L4	AWS g6.xlarge	$0.8048 on-demand	$587.50	$193.52	$1.94	416 estimate	83.1	—	83.1 olmOCR-Bench overall
Chandra OCR	NVIDIA H100 80GB SXM × 8	Google Cloud a3-highgpu-8g	$88.49 on-demand	$64,598	$238.18	$2.38	37,152 estimate	83.1	—	83.1 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA GeForce RTX 4090	Vast.ai RTX 4090 24GB on-demand median	$0.4044 median	$295.24	$20.98	$0.21	1,928 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA A10 / A10G	Vast.ai A10 on-demand median	$0.2453 median	$179.09	$21.38	$0.21	1,148 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA A100 80GB SXM	Vast.ai A100 SXM4 80GB on-demand median	$1.1253 median	$821.49	$28.85	$0.29	3,900 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA H100 80GB SXM	Vast.ai H100 SXM on-demand median	$2.2044 median	$1,609	$34.40	$0.34	6,408 measured	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA A100 80GB SXM	RunPod Secure Cloud A100 SXM 80GB	$1.79 on-demand Community $1.39	$1,307	$45.89	$0.46	3,900 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA GeForce RTX 4090	RunPod Secure Cloud RTX 4090	$0.89 on-demand Community $0.34	$649.70	$46.16	$0.46	1,928 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA A100 80GB PCIe	RunPod Secure Cloud A100 PCIe 80GB	$1.79 on-demand Community $1.19	$1,307	$48.36	$0.48	3,701 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA L4	Vast.ai L4 on-demand median	$0.3378 median	$246.58	$58.86	$0.59	574 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA H100 80GB SXM	Modal Nvidia H100 SXM5	$3.9492 preemptible Non-preemptible $11.85	$2,883	$61.63	$0.62	6,408 measured	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA H100 80GB SXM	RunPod Secure Cloud H100 SXM	$3.99 on-demand Community $2.69	$2,913	$62.27	$0.62	6,408 measured	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA A100 80GB SXM	Modal Nvidia A100 80GB	$2.4984 preemptible Non-preemptible $7.4952	$1,824	$64.06	$0.64	3,900 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA H100 80GB PCIe	Vast.ai H100 PCIe on-demand median	$2.5422 median	$1,856	$66.45	$0.66	3,826 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA A100 40GB SXM	Lambda 1× A100 SXM 40GB	$1.99 on-demand	$1,453	$66.90	$0.67	2,974 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA H100 80GB SXM	Lambda 1× H100 SXM	$4.29 on-demand	$3,132	$66.95	$0.67	6,408 measured	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA GeForce RTX 4090	RunPod Serverless RTX 4090 worker	$1.34 on-demand	$978.20	$69.50	$0.69	1,928 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA A100 80GB SXM	RunPod Serverless A100 80GB worker	$2.72 on-demand	$1,986	$69.74	$0.70	3,900 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA H100 80GB PCIe	RunPod Secure Cloud H100 PCIe	$2.89 on-demand Community $1.99	$2,110	$75.54	$0.76	3,826 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA H100 80GB SXM	RunPod Serverless H100 80GB worker	$5.39 on-demand	$3,935	$84.11	$0.84	6,408 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA H100 80GB PCIe	Lambda 1× H100 PCIe	$3.29 on-demand	$2,402	$86.00	$0.86	3,826 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA A10 / A10G	AWS g5.xlarge	$1.006 on-demand	$734.38	$87.65	$0.88	1,148 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA A100 80GB SXM × 8	AWS p4de.24xlarge	$27.45 on-demand	$20,036	$87.97	$0.88	31,202 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA A10 / A10G	Modal Nvidia A10	$1.1016 preemptible Non-preemptible $3.3048	$804.17	$95.98	$0.96	1,148 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA L4	RunPod Secure Cloud L4	$0.59 on-demand Community $0.44	$430.70	$102.81	$1.03	574 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA H100 80GB SXM	AWS p5.4xlarge	$6.88 on-demand	$5,022	$107.37	$1.07	6,408 measured	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA A10 / A10G	Lambda 1× A10	$1.29 on-demand	$941.70	$112.40	$1.12	1,148 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA L4	Google Cloud g2-standard-4	$0.7068 on-demand	$515.99	$123.17	$1.23	574 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA A100 80GB SXM	Google Cloud a2-ultragpu-1g	$5.0688 on-demand	$3,700	$129.96	$1.30	3,900 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA L4	Modal Nvidia L4	$0.7992 preemptible Non-preemptible $2.3976	$583.42	$139.27	$1.39	574 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA L4	AWS g6.xlarge	$0.8048 on-demand	$587.50	$140.25	$1.40	574 estimate	82.4	—	82.4 olmOCR-Bench overall
olmOCR-2-7B	NVIDIA H100 80GB SXM × 8	Google Cloud a3-highgpu-8g	$88.49 on-demand	$64,598	$172.62	$1.73	51,264 estimate	82.4	—	82.4 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA GeForce RTX 4090	Vast.ai RTX 4090 24GB on-demand median	$0.4044 median	$295.24	$16.97	$0.17	2,383 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA A10 / A10G	Vast.ai A10 on-demand median	$0.2453 median	$179.09	$17.30	$0.17	1,419 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA A100 80GB SXM	Vast.ai A100 SXM4 80GB on-demand median	$1.1253 median	$821.49	$23.34	$0.23	4,821 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA H100 80GB SXM	Vast.ai H100 SXM on-demand median	$2.2044 median	$1,609	$27.83	$0.28	7,920 measured	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA A100 80GB SXM	RunPod Secure Cloud A100 SXM 80GB	$1.79 on-demand Community $1.39	$1,307	$37.13	$0.37	4,821 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA GeForce RTX 4090	RunPod Secure Cloud RTX 4090	$0.89 on-demand Community $0.34	$649.70	$37.35	$0.37	2,383 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA A100 80GB PCIe	RunPod Secure Cloud A100 PCIe 80GB	$1.79 on-demand Community $1.19	$1,307	$39.13	$0.39	4,575 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA L4	Vast.ai L4 on-demand median	$0.3378 median	$246.58	$47.62	$0.48	709 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA H100 80GB SXM	Modal Nvidia H100 SXM5	$3.9492 preemptible Non-preemptible $11.85	$2,883	$49.86	$0.50	7,920 measured	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA H100 80GB SXM	RunPod Secure Cloud H100 SXM	$3.99 on-demand Community $2.69	$2,913	$50.38	$0.50	7,920 measured	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA A100 80GB SXM	Modal Nvidia A100 80GB	$2.4984 preemptible Non-preemptible $7.4952	$1,824	$51.83	$0.52	4,821 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA H100 80GB PCIe	Vast.ai H100 PCIe on-demand median	$2.5422 median	$1,856	$53.77	$0.54	4,728 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA A100 40GB SXM	Lambda 1× A100 SXM 40GB	$1.99 on-demand	$1,453	$54.13	$0.54	3,676 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA H100 80GB SXM	Lambda 1× H100 SXM	$4.29 on-demand	$3,132	$54.17	$0.54	7,920 measured	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA GeForce RTX 4090	RunPod Serverless RTX 4090 worker	$1.34 on-demand	$978.20	$56.23	$0.56	2,383 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA A100 80GB SXM	RunPod Serverless A100 80GB worker	$2.72 on-demand	$1,986	$56.42	$0.56	4,821 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA H100 80GB PCIe	RunPod Secure Cloud H100 PCIe	$2.89 on-demand Community $1.99	$2,110	$61.12	$0.61	4,728 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA H100 80GB SXM	RunPod Serverless H100 80GB worker	$5.39 on-demand	$3,935	$68.06	$0.68	7,920 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA H100 80GB PCIe	Lambda 1× H100 PCIe	$3.29 on-demand	$2,402	$69.58	$0.70	4,728 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA A10 / A10G	AWS g5.xlarge	$1.006 on-demand	$734.38	$70.92	$0.71	1,419 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA A100 80GB SXM × 8	AWS p4de.24xlarge	$27.45 on-demand	$20,036	$71.17	$0.71	38,564 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA A10 / A10G	Modal Nvidia A10	$1.1016 preemptible Non-preemptible $3.3048	$804.17	$77.66	$0.78	1,419 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA L4	RunPod Secure Cloud L4	$0.59 on-demand Community $0.44	$430.70	$83.19	$0.83	709 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA H100 80GB SXM	AWS p5.4xlarge	$6.88 on-demand	$5,022	$86.87	$0.87	7,920 measured	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA A10 / A10G	Lambda 1× A10	$1.29 on-demand	$941.70	$90.94	$0.91	1,419 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA L4	Google Cloud g2-standard-4	$0.7068 on-demand	$515.99	$99.66	$1.00	709 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA A100 80GB SXM	Google Cloud a2-ultragpu-1g	$5.0688 on-demand	$3,700	$105.15	$1.05	4,821 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA L4	Modal Nvidia L4	$0.7992 preemptible Non-preemptible $2.3976	$583.42	$112.68	$1.13	709 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA L4	AWS g6.xlarge	$0.8048 on-demand	$587.50	$113.47	$1.13	709 estimate	80	94.18	80 olmOCR-Bench overall
PaddleOCR-VL · China	NVIDIA H100 80GB SXM × 8	Google Cloud a3-highgpu-8g	$88.49 on-demand	$64,598	$139.66	$1.40	63,360 estimate	80	94.18	80 olmOCR-Bench overall
dots.ocr · China	NVIDIA GeForce RTX 4090	Vast.ai RTX 4090 24GB on-demand median	$0.4044 median	$295.24	$19.25	$0.19	2,101 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA A10 / A10G	Vast.ai A10 on-demand median	$0.2453 median	$179.09	$19.61	$0.20	1,251 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA A100 80GB SXM	Vast.ai A100 SXM4 80GB on-demand median	$1.1253 median	$821.49	$26.47	$0.26	4,251 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA H100 80GB SXM	Vast.ai H100 SXM on-demand median	$2.2044 median	$1,609	$31.56	$0.32	6,984 measured	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA A100 80GB SXM	RunPod Secure Cloud A100 SXM 80GB	$1.79 on-demand Community $1.39	$1,307	$42.11	$0.42	4,251 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA GeForce RTX 4090	RunPod Secure Cloud RTX 4090	$0.89 on-demand Community $0.34	$649.70	$42.35	$0.42	2,101 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA A100 80GB PCIe	RunPod Secure Cloud A100 PCIe 80GB	$1.79 on-demand Community $1.19	$1,307	$44.37	$0.44	4,034 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA L4	Vast.ai L4 on-demand median	$0.3378 median	$246.58	$54.01	$0.54	625 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA H100 80GB SXM	Modal Nvidia H100 SXM5	$3.9492 preemptible Non-preemptible $11.85	$2,883	$56.55	$0.57	6,984 measured	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA H100 80GB SXM	RunPod Secure Cloud H100 SXM	$3.99 on-demand Community $2.69	$2,913	$57.13	$0.57	6,984 measured	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA A100 80GB SXM	Modal Nvidia A100 80GB	$2.4984 preemptible Non-preemptible $7.4952	$1,824	$58.77	$0.59	4,251 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA H100 80GB PCIe	Vast.ai H100 PCIe on-demand median	$2.5422 median	$1,856	$60.97	$0.61	4,170 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA A100 40GB SXM	Lambda 1× A100 SXM 40GB	$1.99 on-demand	$1,453	$61.39	$0.61	3,242 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA H100 80GB SXM	Lambda 1× H100 SXM	$4.29 on-demand	$3,132	$61.43	$0.61	6,984 measured	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA GeForce RTX 4090	RunPod Serverless RTX 4090 worker	$1.34 on-demand	$978.20	$63.77	$0.64	2,101 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA A100 80GB SXM	RunPod Serverless A100 80GB worker	$2.72 on-demand	$1,986	$63.99	$0.64	4,251 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA H100 80GB PCIe	RunPod Secure Cloud H100 PCIe	$2.89 on-demand Community $1.99	$2,110	$69.31	$0.69	4,170 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA H100 80GB SXM	RunPod Serverless H100 80GB worker	$5.39 on-demand	$3,935	$77.18	$0.77	6,984 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA H100 80GB PCIe	Lambda 1× H100 PCIe	$3.29 on-demand	$2,402	$78.91	$0.79	4,170 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA A10 / A10G	AWS g5.xlarge	$1.006 on-demand	$734.38	$80.42	$0.80	1,251 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA A100 80GB SXM × 8	AWS p4de.24xlarge	$27.45 on-demand	$20,036	$80.71	$0.81	34,007 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA A10 / A10G	Modal Nvidia A10	$1.1016 preemptible Non-preemptible $3.3048	$804.17	$88.07	$0.88	1,251 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA L4	RunPod Secure Cloud L4	$0.59 on-demand Community $0.44	$430.70	$94.33	$0.94	625 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA H100 80GB SXM	AWS p5.4xlarge	$6.88 on-demand	$5,022	$98.51	$0.99	6,984 measured	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA A10 / A10G	Lambda 1× A10	$1.29 on-demand	$941.70	$103.13	$1.03	1,251 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA L4	Google Cloud g2-standard-4	$0.7068 on-demand	$515.99	$113.01	$1.13	625 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA A100 80GB SXM	Google Cloud a2-ultragpu-1g	$5.0688 on-demand	$3,700	$119.24	$1.19	4,251 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA L4	Modal Nvidia L4	$0.7992 preemptible Non-preemptible $2.3976	$583.42	$127.78	$1.28	625 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA L4	AWS g6.xlarge	$0.8048 on-demand	$587.50	$128.68	$1.29	625 estimate	79.1	90.77	79.1 olmOCR-Bench overall
dots.ocr · China	NVIDIA H100 80GB SXM × 8	Google Cloud a3-highgpu-8g	$88.49 on-demand	$64,598	$158.38	$1.58	55,872 estimate	79.1	90.77	79.1 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA GeForce RTX 4090	Vast.ai RTX 4090 24GB on-demand median	$0.4044 median	$295.24	$8.03	$0.0803	5,037 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA A10 / A10G	Vast.ai A10 on-demand median	$0.2453 median	$179.09	$8.18	$0.0818	2,998 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA A100 80GB SXM	Vast.ai A100 SXM4 80GB on-demand median	$1.1253 median	$821.49	$11.04	$0.11	10,189 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA H100 80GB SXM	Vast.ai H100 SXM on-demand median	$2.2044 median	$1,609	$13.17	$0.13	16,740 measured	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA A100 80GB SXM	RunPod Secure Cloud A100 SXM 80GB	$1.79 on-demand Community $1.39	$1,307	$17.57	$0.18	10,189 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA GeForce RTX 4090	RunPod Secure Cloud RTX 4090	$0.89 on-demand Community $0.34	$649.70	$17.67	$0.18	5,037 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA A100 80GB PCIe	RunPod Secure Cloud A100 PCIe 80GB	$1.79 on-demand Community $1.19	$1,307	$18.51	$0.19	9,669 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA L4	Vast.ai L4 on-demand median	$0.3378 median	$246.58	$22.53	$0.23	1,499 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA H100 80GB SXM	Modal Nvidia H100 SXM5	$3.9492 preemptible Non-preemptible $11.85	$2,883	$23.59	$0.24	16,740 measured	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA H100 80GB SXM	RunPod Secure Cloud H100 SXM	$3.99 on-demand Community $2.69	$2,913	$23.84	$0.24	16,740 measured	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA A100 80GB SXM	Modal Nvidia A100 80GB	$2.4984 preemptible Non-preemptible $7.4952	$1,824	$24.52	$0.25	10,189 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA H100 80GB PCIe	Vast.ai H100 PCIe on-demand median	$2.5422 median	$1,856	$25.44	$0.25	9,994 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA A100 40GB SXM	Lambda 1× A100 SXM 40GB	$1.99 on-demand	$1,453	$25.61	$0.26	7,770 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA H100 80GB SXM	Lambda 1× H100 SXM	$4.29 on-demand	$3,132	$25.63	$0.26	16,740 measured	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA GeForce RTX 4090	RunPod Serverless RTX 4090 worker	$1.34 on-demand	$978.20	$26.60	$0.27	5,037 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA A100 80GB SXM	RunPod Serverless A100 80GB worker	$2.72 on-demand	$1,986	$26.70	$0.27	10,189 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA H100 80GB PCIe	RunPod Secure Cloud H100 PCIe	$2.89 on-demand Community $1.99	$2,110	$28.92	$0.29	9,994 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA H100 80GB SXM	RunPod Serverless H100 80GB worker	$5.39 on-demand	$3,935	$32.20	$0.32	16,740 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA H100 80GB PCIe	Lambda 1× H100 PCIe	$3.29 on-demand	$2,402	$32.92	$0.33	9,994 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA A10 / A10G	AWS g5.xlarge	$1.006 on-demand	$734.38	$33.55	$0.34	2,998 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA A100 80GB SXM × 8	AWS p4de.24xlarge	$27.45 on-demand	$20,036	$33.67	$0.34	81,511 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA A10 / A10G	Modal Nvidia A10	$1.1016 preemptible Non-preemptible $3.3048	$804.17	$36.74	$0.37	2,998 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA L4	RunPod Secure Cloud L4	$0.59 on-demand Community $0.44	$430.70	$39.36	$0.39	1,499 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA H100 80GB SXM	AWS p5.4xlarge	$6.88 on-demand	$5,022	$41.10	$0.41	16,740 measured	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA A10 / A10G	Lambda 1× A10	$1.29 on-demand	$941.70	$43.03	$0.43	2,998 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA L4	Google Cloud g2-standard-4	$0.7068 on-demand	$515.99	$47.15	$0.47	1,499 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA A100 80GB SXM	Google Cloud a2-ultragpu-1g	$5.0688 on-demand	$3,700	$49.75	$0.50	10,189 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA L4	Modal Nvidia L4	$0.7992 preemptible Non-preemptible $2.3976	$583.42	$53.31	$0.53	1,499 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA L4	AWS g6.xlarge	$0.8048 on-demand	$587.50	$53.69	$0.54	1,499 estimate	75.7	—	75.7 olmOCR-Bench overall
DeepSeek-OCR · China	NVIDIA H100 80GB SXM × 8	Google Cloud a3-highgpu-8g	$88.49 on-demand	$64,598	$66.08	$0.66	133,920 estimate	75.7	—	75.7 olmOCR-Bench overall
Textract	—	Managed API Detect Document Text	—	—	$150.00	$1.50	—	not benchmarked	not benchmarked	—
Document AI	—	Managed API Enterprise Document OCR	—	—	$150.00	$1.50	—	not benchmarked	not benchmarked	—
Azure	—	Managed API S0 Read	—	—	$150.00	$1.50	—	not benchmarked	not benchmarked	—
Mistral OCR	—	Managed API OCR 4.1	—	—	$400.00	$4.00	—	72 (not OCR 4.1)	85.66	—

“—” means that score is not in the sources checked on 2026-11-01. “Not benchmarked” means the managed API has no row on that board. Mistral’s 72.0 is the olmOCR-Bench README row named “Mistral OCR API”, and 85.66 is the OmniDocBench row named “Mistral OCR”. Neither row says OCR 4.1, which is the price in the cost columns.

Chinese models with a score, and no monthly cost

These rows are on the OmniDocBench v1.6 leaderboard, read 2026-11-01. They were not in the E2E H100 pages-per-second table, so this page does not invent a GPU bill for them. PaddleOCR-VL, dots.ocr, and DeepSeek-OCR are priced above because that H100 run measured those three.

OmniDocBench v1.6 scores for models with no measured GPU throughput, read 2026-11-01.
Model	Lab	Size	OmniDocBench	Table TEDS	Formula CDM	Text edit
PaddleOCR-VL-1.6	Baidu	0.9B	96.34	94.7619	97.5304	0.0326
MinerU2.5-Pro	MinerU	1.2B	95.75	93.42	97.45	0.036
GLM-OCR	GLM	0.9B	95.22	92.83	97.18	0.044
PaddleOCR-VL-1.5	Baidu	0.9B	94.93	91.67	96.89	0.038
MinerU2.5	MinerU	1.2B	93.04	87.88	95.77	0.045
DeepSeek-OCR 2	DeepSeek	3B	90.25	83.89	91.84	0.05
HunyuanOCR	Hunyuan	1B	89.95	91.01	87.68	0.088
Qwen3-VL-235B	Qwen	235B	89.78	83.07	92.55	0.063

Overall is higher-better. Text edit is lower-better and is not turned into a 0–100 score. Qwen3-VL-235B is a general model, not a single-GPU OCR SKU in the price grid. DeepSeek-OCR 2 is not the DeepSeek-OCR row in the cost table.

OCR GPU cost per hour and per month, by provider

On-demand hourly price and the always-on month (hourly × 730). Community and non-preemptible rates are shown only where that provider publishes them. Checked 2026-11-01. A dash means that provider does not list this GPU.

On-demand GPU price per hour and per 730-hour month, by provider. Checked 2026-11-01.
GPU	RunPod	Lambda	Modal	Vast.ai	AWS	Google Cloud
L4	L4 $0.59/hr On-demand $430.70/mo Community $0.44/hr · $321.20/mo	—	Nvidia L4 $0.7992/hr Preemptible $583.42/mo Non-preemptible $2.3976/hr · $1,750/mo	L4 on-demand median $0.3378/hr On-demand median $246.58/mo	g6.xlarge $0.8048/hr On-demand $587.50/mo	g2-standard-4 $0.7068/hr On-demand $515.99/mo
A10 / A10G	—	1× A10 $1.29/hr On-demand $941.70/mo	Nvidia A10 $1.1016/hr Preemptible $804.17/mo Non-preemptible $3.3048/hr · $2,413/mo	A10 on-demand median $0.2453/hr On-demand median $179.09/mo	g5.xlarge $1.006/hr On-demand $734.38/mo	—
RTX 4090	RTX 4090 $0.89/hr On-demand $649.70/mo Community $0.34/hr · $248.20/mo	—	—	RTX 4090 24GB on-demand median $0.4044/hr On-demand median $295.24/mo	—	—
A100	A100 SXM 80GB $1.79/hr On-demand $1,307/mo Community $1.39/hr · $1,015/mo A100 PCIe 80GB $1.79/hr On-demand $1,307/mo Community $1.19/hr · $868.70/mo	1× A100 SXM 40GB $1.99/hr On-demand $1,453/mo	Nvidia A100 80GB $2.4984/hr Preemptible $1,824/mo Non-preemptible $7.4952/hr · $5,471/mo	A100 SXM4 80GB on-demand median $1.1253/hr On-demand median $821.49/mo	p4de.24xlarge $27.45/hr On-demand · 8 GPUs $20,036/mo	a2-ultragpu-1g $5.0688/hr On-demand $3,700/mo
H100	H100 SXM $3.99/hr On-demand $2,913/mo Community $2.69/hr · $1,964/mo H100 PCIe $2.89/hr On-demand $2,110/mo Community $1.99/hr · $1,453/mo	1× H100 SXM $4.29/hr On-demand $3,132/mo 1× H100 PCIe $3.29/hr On-demand $2,402/mo	Nvidia H100 SXM5 $3.9492/hr Preemptible $2,883/mo Non-preemptible $11.85/hr · $8,649/mo	H100 SXM on-demand median $2.2044/hr On-demand median $1,609/mo H100 PCIe on-demand median $2.5422/hr On-demand median $1,856/mo	p5.4xlarge $6.88/hr On-demand $5,022/mo	a3-highgpu-8g $88.49/hr On-demand · 8 GPUs $64,598/mo

GPU cost per 1,000 pages: the cheapest setup per model

Cheapest GPU setup for each open OCR model at 100,000 pages a month, with the monthly bill and the cost per 1,000 pages.
Model	Setup	Pages/hour	This month	Per 1,000 pages	Score
olmOCR-2-7B	Vast.ai RTX 4090 24GB on-demand median	1,928 est.	$20.98	$0.21	82.4
PaddleOCR-VL	Vast.ai RTX 4090 24GB on-demand median	2,383 est.	$16.97	$0.17	80
dots.ocr	Vast.ai RTX 4090 24GB on-demand median	2,101 est.	$19.25	$0.19	79.1
Chandra OCR	Vast.ai RTX 4090 24GB on-demand median	1,397 est.	$28.94	$0.29	83.1
DeepSeek-OCR	Vast.ai RTX 4090 24GB on-demand median	5,037 est.	$8.03	$0.0803	75.7

How the self-hosted OCR cost is calculated

Prices checked 2026-11-01. A month is 730 hours. The bill is GPU rental only. Quality is a property of the model, not the GPU. The same model has the same score on an L4 and an H100. The GPU changes pages per hour and the bill. The recommendation is the highest published score for the document types you picked, among models that cost under twice the cheapest GPU-hour bill. That 2× cap is a product rule, not a benchmark. Clean PDFs use the overall olmOCR-Bench score. Scans use that same overall score, because no shared scanned-page score is published for every model here. Tables prefer OmniDocBench table TEDS when a model has one. Handwriting has no published score for these models, so it also uses overall olmOCR-Bench and does not pretend one model won a handwriting test. You can select more than one type. The page count stays one number and is not split, because no share of pages was entered. If tables are included, that score sets the rank, and the API price uses the tables SKU for the full month. Quality does not change with the GPU.

Sources, bandwidth scale, and what was left out

Non-H100 rates are the E2E Networks H100 pages/sec multiplied by the ratio of published NVIDIA memory bandwidth to the H100 SXM figure of 3.35 TB/s. Small vision-language OCR models are usually bandwidth-bound, which is why bandwidth is the scale, not a TFLOPS spec. E2E did not say whether their H100 was SXM or PCIe. This uses the SXM bandwidth as the denominator. A multi-GPU instance multiplies that per-GPU rate by the GPU count. That linear step is also an estimate. A 24 GB card cannot run the batch sizes some papers used on an 80 GB GPU, so a scaled 24 GB rate can run high.

Models and the H100 measurement

olmOCR-2-7B — 1.78 pages/sec on one H100, olmOCR-Bench score 82.4. Allen Institute for AI. Fine-tune of Qwen2.5-VL-7B. FP8 weights are published for a single H100. 82.4 on olmOCR-Bench in the E2E table (±1.1 there). Ai2's 22 October 2025 release note states 82.4, and gives 84.9 on tables, 82.3 on old math scans, and 83.7 on multi-column. The project paper's table lists 82.3 ± 0.9 overall for the November 2025 release. The OmniDocBench leaderboard row named olmOCR 7B is not labeled olmOCR-2, so it is not used as this model's OmniDocBench score. E2E Networks, open-source OCR benchmark on one H100. Ai2 also reports 3,400 output tokens/sec on one H100 for the FP8 model, and "10,000 pages for less than $2". That is not converted into pages/hour here, because the post does not state tokens per page.
PaddleOCR-VL — 2.2 pages/sec on one H100, olmOCR-Bench score 80. Baidu PaddlePaddle 0.9B vision-language document parser. The priced row is this model, the one in the E2E H100 table. PaddleOCR-VL-1.5 and 1.6 are newer leaderboard rows and are not given this throughput. 80.0 ± 1.0 on olmOCR-Bench. The PaddleOCR-VL paper's Table 4 also gives ArXiv 85.7, headers and footers 97.0, multi-column 79.9, and long tiny text 85.7. OmniDocBench is a different scale and is listed separately. E2E Networks, open-source OCR benchmark on one H100. The PaddleOCR-VL paper (arXiv 2510.14528, Table 13) measured 1.2241 pages/s on one A100 with vLLM, batch 512, about 43.7 GB of VRAM, on OmniDocBench v1.0. A bandwidth scale of the E2E H100 rate down to an A100 80GB SXM lands near that (see the methodology). The paper's batch does not fit a 24 GB GPU. Paper
dots.ocr — 1.94 pages/sec on one H100, olmOCR-Bench score 79.1. RED AI Lab (rednote-hilab/dots.ocr). E2E describes a 3B fine-tune of Qwen2.5-VL. Included because it sits next to olmOCR-2 and PaddleOCR-VL on the same H100 measurement. 79.1 ± 1.0 on olmOCR-Bench in the E2E table. No per-category olmOCR-Bench row for dots.ocr was on the bench README read 1 November 2026. E2E Networks, open-source OCR benchmark on one H100. The PaddleOCR-VL paper measured dots.ocr at 0.3522 pages/s on the same A100 vLLM setup (Table 13). That is much slower than a bandwidth scale of the E2E H100 rate. The scaled dots.ocr rows can be optimistic. The H100 row is the E2E measurement, not a scale. Paper
Chandra OCR — 1.29 pages/sec on one H100, olmOCR-Bench score 83.1. Highest olmOCR-Bench score in the E2E table. Slower than the other three on that same H100 run. 83.1 ± 0.9 on olmOCR-Bench in the E2E table. The olmOCR project paper reports the same 83.1 ± 0.9 for Chandra OCR 0.1.0. E2E Networks, open-source OCR benchmark on one H100.
DeepSeek-OCR — 4.65 pages/sec on one H100, olmOCR-Bench score 75.7. DeepSeek-AI. About 3B parameters, with a mixture-of-experts decoder E2E describes as roughly 570M active. The OmniDocBench v1.6 row named DeepSeek-OCR 2 is a later model and is not given this throughput. 75.7 ± 1.0 on olmOCR-Bench in the E2E H100 table, 4.65 pages/sec on that same run. The v1.6 leaderboard read on 1 November 2026 does not list this model. It lists DeepSeek-OCR 2, so no OmniDocBench score is attached here. E2E Networks, open-source OCR benchmark on one H100.

GPU bandwidth used for estimates

NVIDIA L4: 300 GB/s. NVIDIA L4 product specifications, 300 GB/s
NVIDIA A10 / A10G: 600 GB/s. NVIDIA A10 specifications, 600 GB/s AWS calls this GPU the A10G on g5 instances. The bandwidth used is the A10's published 600 GB/s.
NVIDIA GeForce RTX 4090: 1,008 GB/s. NVIDIA Ada GPU architecture whitepaper, GeForce RTX 4090, 1008 GB/s
NVIDIA A100 80GB SXM: 2,039 GB/s. NVIDIA A100 specifications
NVIDIA A100 80GB PCIe: 1,935 GB/s. NVIDIA A100 specifications
NVIDIA A100 40GB SXM: 1,555 GB/s. NVIDIA A100 datasheet, 40 GB bandwidth 1,555 GB/s
NVIDIA H100 80GB SXM: 3,350 GB/s. NVIDIA H100 datasheet, SXM memory bandwidth 3.35 TB/s
NVIDIA H100 80GB PCIe: 2,000 GB/s. NVIDIA H100 PCIe product brief, peak memory bandwidth 2,000 GB/s

GPU prices checked 2026-11-01

RunPod Secure Cloud L4 — $0.59/hr. RunPod Secure Cloud. Secure Cloud is the rate used. Community Cloud is also listed at $0.44/hr. RunPod pricing
RunPod Secure Cloud RTX 4090 — $0.89/hr. RunPod Secure Cloud. Secure Cloud is the rate used. Community Cloud is also listed at $0.34/hr. RunPod pricing
RunPod Secure Cloud A100 SXM 80GB — $1.79/hr. RunPod Secure Cloud. The pricing card says 80 GB VRAM. Community Cloud is $1.39/hr. RunPod pricing
RunPod Secure Cloud A100 PCIe 80GB — $1.79/hr. RunPod Secure Cloud. The pricing card says 80 GB VRAM. Community Cloud is $1.19/hr. RunPod pricing
RunPod Secure Cloud H100 SXM — $3.99/hr. RunPod Secure Cloud. Community Cloud is $2.69/hr. RunPod pricing
RunPod Secure Cloud H100 PCIe — $2.89/hr. RunPod Secure Cloud. Community Cloud is $1.99/hr. RunPod pricing
RunPod Serverless RTX 4090 worker — $1.34/hr. RunPod Serverless. Serverless price is per hour while a worker runs. The 24 GB class that mixes L4 with other GPUs is not used, because that card is not an L4-only price. RunPod pricing
RunPod Serverless A100 80GB worker — $2.72/hr. RunPod Serverless. The serverless card says A100, 80 GB. It does not say SXM or PCIe. Throughput uses SXM bandwidth and is marked an estimate. RunPod pricing
RunPod Serverless H100 80GB worker — $5.39/hr. RunPod Serverless. The serverless card says H100, 80 GB. It does not say SXM or PCIe. Throughput uses SXM bandwidth and is marked an estimate. RunPod pricing
Lambda 1× A10 — $1.29/hr. Lambda 1× instance. 1× row. Price is per GPU-hour. Lambda says per minute, no egress fees, plus tax. Lambda GPU Cloud, 1× instance prices
Lambda 1× A100 SXM 40GB — $1.99/hr. Lambda 1× instance. The 1× A100 on the page is 40 GB SXM, not 80 GB. The 8× A100 80GB row is $2.79 per GPU and is a different machine, so it is not used as a single-GPU price. Lambda GPU Cloud, 1× instance prices
Lambda 1× H100 SXM — $4.29/hr. Lambda 1× instance. Lambda GPU Cloud, 1× instance prices
Lambda 1× H100 PCIe — $3.29/hr. Lambda 1× instance. Lambda GPU Cloud, 1× instance prices
Modal Nvidia L4 — $0.7992/hr. Modal Functions. List rate is the preemptible Function price, $0.000222/sec. Non-preemptible is 3× on Modal's page. Real-time uses the 3× rate and assumes the container stays up. CPU, memory, and the 60-second idle window are extra and not added. Modal pricing, per-second GPU rates
Modal Nvidia A10 — $1.1016/hr. Modal Functions. Preemptible list rate $0.000306/sec. Real-time uses the published 3× non-preemptible multiplier. Modal pricing, per-second GPU rates
Modal Nvidia A100 80GB — $2.4984/hr. Modal Functions. Preemptible list rate $0.000694/sec. Modal says 80 GB and does not say SXM or PCIe. Throughput uses SXM bandwidth and is marked an estimate. Modal pricing, per-second GPU rates
Modal Nvidia H100 SXM5 — $3.9492/hr. Modal Functions. The page names H100 SXM5 at $0.001097/sec. Real-time uses the 3× non-preemptible multiplier. Starter includes $30 of compute a month. That credit is not subtracted. Modal pricing, per-second GPU rates
Vast.ai L4 on-demand median — $0.3378/hr. Marketplace. Median of 20 on-demand L4 offers at 01:12 UTC. The low was $0.096 and the high was $0.818. dph_total includes 16 GB of disk. This is not a list price. Vast.ai public bundles API, on-demand, 1 GPU, 16 GB disk
Vast.ai A10 on-demand median — $0.2453/hr. Marketplace. Only 4 on-demand A10 offers came back, all at this rate. Treat it as a thin sample. Vast.ai public bundles API, on-demand, 1 GPU, 16 GB disk
Vast.ai RTX 4090 24GB on-demand median — $0.4044/hr. Marketplace. Median of 60 offers reporting about 24 GB. Four offers reporting about 48 GB were left out. Low $0.123, high $0.471. Vast.ai public bundles API, on-demand, 1 GPU, 16 GB disk
Vast.ai A100 SXM4 80GB on-demand median — $1.1253/hr. Marketplace. Median of 18 A100 SXM4 offers reporting 80 GB. 40 GB offers were not mixed in. Vast.ai public bundles API, on-demand, 1 GPU, 16 GB disk
Vast.ai H100 SXM on-demand median — $2.2044/hr. Marketplace. Median of 22 on-demand H100 SXM offers reporting about 80 GB. Vast.ai public bundles API, on-demand, 1 GPU, 16 GB disk
Vast.ai H100 PCIe on-demand median — $2.5422/hr. Marketplace. Median of 6 on-demand H100 PCIe offers. Small sample. Vast.ai public bundles API, on-demand, 1 GPU, 16 GB disk
AWS g6.xlarge — $0.8048/hr. us-east-1. 1× L4, 4 vCPU, 16 GiB. Linux on-demand. EBS, data transfer, and tax are not included. EC2 bills per second after the first minute. AWS EC2 on-demand, US East (N. Virginia), Linux Read from the price file that page loads: b0.p.awsstatic.com/.../US East (N. Virginia)/Linux/index.json on 1 November 2026.
AWS g5.xlarge — $1.006/hr. us-east-1. 1× A10G, 4 vCPU, 16 GiB. Linux on-demand. AWS EC2 on-demand, US East (N. Virginia), Linux Read from the price file that page loads: b0.p.awsstatic.com/.../US East (N. Virginia)/Linux/index.json on 1 November 2026.
AWS p5.4xlarge — $6.88/hr. us-east-1. 1× H100. p5.48xlarge is $55.04/hr for 8 GPUs, the same $6.88 per GPU. The single-GPU size is what this row uses. AWS EC2 on-demand, US East (N. Virginia), Linux Read from the price file that page loads: b0.p.awsstatic.com/.../US East (N. Virginia)/Linux/index.json on 1 November 2026.
AWS p4de.24xlarge — $27.45/hr for 8 GPUs. us-east-1. 8× A100 80GB. You rent the whole instance at $27.44705/hr. There is no 1× A100 80GB on-demand size in this price file. p4d.24xlarge is the 40GB A100 at $21.957642/hr for 8 GPUs and is not listed separately. AWS EC2 on-demand, US East (N. Virginia), Linux Read from the price file that page loads: b0.p.awsstatic.com/.../US East (N. Virginia)/Linux/index.json on 1 November 2026.
Google Cloud g2-standard-4 — $0.7068/hr. us-central1. 1× L4, 4 vCPU, 16 GiB, on-demand in Iowa. Spot and commitment columns on the same row are not used. Google Cloud accelerator-optimized VM pricing, Iowa (us-central1)
Google Cloud a2-ultragpu-1g — $5.0688/hr. us-central1. On-demand hourly price for a2-ultragpu-1g in Iowa. Google's GPU docs describe A2 Ultra as A100 80GB. a2-highgpu-1g is the 40GB A100 at $3.673385/hr and is not the row used here. Google Cloud accelerator-optimized VM pricing, Iowa (us-central1)
Google Cloud a3-highgpu-8g — $88.49/hr for 8 GPUs. us-central1. 8× H100. The on-demand price is for the whole VM. Google does not sell a 1× H100 on-demand from this table. Google Cloud accelerator-optimized VM pricing, Iowa (us-central1)

API prices checked 2026-11-01

AWS Textract — text: $0.0015 per page for the first 1M pages/month, then $0.0006. Structured: $0.015 per page. Taken from AWS's own pricing examples for US West (Oregon). The page's examples use $0.0015 and $0.0006. AWS's forms-and-tables example prices a page with tables at $0.015. Forms (key-value) are $0.05 per page in that same example and are not the rate used here. A volume discount for Tables was not taken from a third-party copy of the price list. AWS Textract pricing
Google Document AI — text: $1.50 per 1,000 pages from 1,000 to 5,000,000 pages/month, then $0.60 per 1,000. Structured: $10 per 1,000 pages. The first 1,000 pages in a month are free. That free block is not subtracted. Layout Parser, including initial chunking, on the same pricing page. No volume break was listed for it. Google Document AI pricing
Azure Document Intelligence — text: $1.50 per 1,000 pages, then $0.60 per 1,000 after 1M pages. Structured: $10 per 1,000 pages. The marketing page draws prices in the browser, so the dollars are from Microsoft's retail price API for East US: meter "S0 Read Pages", $1.50 at tierMinimumUnits 0 and $0.60 at tierMinimumUnits 1000 (the unit is 1,000 pages), effective 2022-03-01. API: https://prices.azure.com/api/retail/prices Retail meter "S0 Pre-built Pages", East US, $10 per 1K, effective 2021-08-01. Microsoft's page groups Layout with the other prebuilt models on this meter. Azure Document Intelligence pricing
Mistral OCR — text: $4 per 1,000 pages. Structured: $4 per 1,000 pages. docs.mistral.ai lists OCR 4.1 input at $4 / 1,000 pages. The model page states the same $4 / 1,000 pages, and $5 / 1,000 annotated pages. Annotated output is not the rate used. A June 2026 Mistral post says the batch API is half price ($2 / 1,000). That discount is not applied. Mistral does not publish a separate tables SKU. The structured comparison uses the same $4 / 1,000 list price. Mistral OCR pricing

Left out on purpose

olmOCR v1 at $176 per million pages. Table 6 of the February 2025 olmOCR paper: $176 per million pages on an L40S, using RunPod's then-current $0.79/hr. That is v1, with a measured 12% retry rate. olmOCR-2 is a different model. The $176 figure is not an input. Source
Hetzner GEX GPU servers. Hetzner does sell GPUs. On 1 November 2026 the GEX line was an RTX PRO 4000 Blackwell (GEX45, €214/month or €0.3429/hour, plus a €209 setup fee, prices exclude VAT) and larger RTX PRO 6000 Blackwell servers. Those are not L4, A10, RTX 4090, A100, or H100 cards. No published OCR pages/hour for the GEX45 was found, so Hetzner is not in the cost math. Source
Nanonets OCR 2. E2E measured 1.12 pages/sec and 69.5 ± 1.1 on olmOCR-Bench on the same H100. That score sits well below the models in this calculator, so it is not offered as a recommendation. The number is theirs, not reused in the math. Source
LightOn OCR. E2E measured 5.55 pages/sec and 76.1 on the same H100 run. It is not a Chinese OCR model, and it is not given a row. The number is theirs, not reused in the math. Source

Deployment checklist

Pin the model commit and the GPU SKU, then time a few hundred of your own pages. The non-H100 rates here are bandwidth estimates.
Decide batch versus a warm GPU from the break-even, not from the hourly price alone.
Keep the PDFs in storage you control. A shared community GPU is the wrong place for a contract that forbids that.
Log pages, GPU-hours, and retries weekly. The v1 paper's 12% retry rate is not applied to olmOCR-2, because v2's retry rate was not in these sources.
Add the engineering month. GPU rental is the only dollar figure in this calculator.

The long version of this arithmetic is the post Self-hosted OCR cost in 2026. Related reading: LLM API pricing, the ERP operations agent (documents in, a person approves the write), LLM cost optimization, and AI agent development. Other calculators: embedding cost, speech-to-text cost, and open-source LLM hosting cost.

Self-hosted OCR cost: questions

What is a self-hosted OCR cost, really?

In this calculator it is GPU rental only: the hourly price times the hours the GPU has to run, or times 730 hours if you leave it on. Engineering time, object storage, egress, and retries are not in the number. A deployment quote is a separate conversation.

When does an always-on AWS L4 beat Textract?

The break-even page count does not depend on the OCR model. AWS g6.xlarge is $0.8048 an hour in us-east-1. Over 730 hours that is $587.50. Detect Document Text is $0.0015 a page for the first million pages. $587.50 divided by $0.0015 is 391,669 pages in a month. The model decides a different question: whether one L4 can finish that many pages. At the bandwidth-scaled rates here, PaddleOCR-VL and olmOCR-2 clear it, and those rates are marked as estimates. Published roundups that say about 392,000 pages are this same g6.xlarge price, rounded.

Why are most throughputs marked estimate?

The only measured pages-per-second figures in this tool are the E2E Networks H100 runs on olmOCR-Bench at 200 DPI. Every other GPU is that rate times a NVIDIA memory-bandwidth ratio. A 24 GB card also cannot run the 40 GB-plus batches some papers used, so those estimates can run high.

Which open model should I use for handwriting?

No handwriting score is published for these models, so handwriting uses overall olmOCR-Bench and does not lock the pick to olmOCR-2. Ai2's release note says the October 2025 mix added 20,000 handwritten and typewritten pages. That is not a score. You can select more than one document type. The page count is not divided between them.

Are the Vast.ai prices list prices?

No. Vast.ai is a marketplace. The rates are medians of on-demand offers returned by their public bundles API at 01:12 UTC on 1 November 2026, with 16 GB of disk included in dph_total. They move. The result says when a median is being used, and how many offers were in the sample.

Does a faster GPU make the OCR more accurate?

No. The published scores belong to the model. An L4 and an H100 running the same model get the same olmOCR-Bench score in this calculator. The GPU changes pages per hour and the monthly bill.

Does the Mistral price include the batch discount?

No. The comparison uses Mistral's list price of $4 per 1,000 pages for OCR 4.1. Mistral's June 2026 announcement says the batch API is $2 per 1,000. Annotated pages are $5 per 1,000. Neither adjustment is applied.