Computer Vision Development Cost in 2026: Build Tiers, Labelling and Cost per Image

Computer vision development cost by tier

Computer vision projects vary more in price than most software, because the same request ("detect damaged parcels") can be a two-week integration with an existing model or a six-month data project. The three tiers below are our published price bands, with what usually sits in each.

If you are not sure which tier applies, a discovery phase ($3k to $8k) is the cheapest way to find out: we test two or three approaches on a sample of your own images and come back with measured accuracy and a fixed estimate. Our computer vision development page describes how those projects run.

What drives the price

Labelling: the cost most estimates leave out

A custom model needs labelled examples: boxes drawn around each object, or masks traced over each region, checked by a second person. Labelling cost is images times objects per image times seconds per label, plus review. Bounding boxes are quick; polygons and pixel masks take many times longer. We do not quote a generic per-image figure because it varies too much with the task. Instead, label 100 of your own images in a free tool such as CVAT or Label Studio, time it, and multiply.

• Start small: a few hundred well-chosen images of the hard cases teach a model more than thousands of easy ones.
• Pre-label with a model: let an existing detector or a vision-language model draw first-pass boxes and have people correct them, which is faster than drawing from scratch.
• Label what the model gets wrong: after the first training run, send only low-confidence and wrong predictions back for labelling.
• Write a labelling guide with examples of edge cases (half-visible objects, reflections, overlapping items) before anyone starts; inconsistent labels cap accuracy.
• Keep a separate test set that is never used for training, labelled with extra care. It is the only honest measure of progress.

Cloud vision API prices (checked October 2026)

Sources: AWS Rekognition pricing and Google Cloud Vision pricing, checked October 2026. Note the Custom Labels model: at $4 an hour, a model left running all month costs about $2,900 before it sees a single image, so it suits batch jobs that start and stop the model, not always-on use.

Three ways to run a vision model: cost per 1,000 images

The vision-language model rows assume about 1,500 input tokens per image (a 1,000 by 1,000 pixel image is 1,296 tokens on Claude, plus a short prompt) and 150 output tokens, at list prices from Claude pricing and Gemini API pricing. GPU prices are AWS on-demand rates for US East from the EC2 on-demand pricing page, checked October 2026, times 730 hours. For a self-hosted model, cost per image is the hourly price divided by the images the machine actually processes in an hour, so measure your own throughput: if a detector sustains 10 images a second on a T4, that is 36,000 an hour and about $0.015 per 1,000 images, a fraction of any API.

The break-even is volume. At 100,000 images a month, a managed API costs about $100 and a dedicated GPU about $384, so the API wins. At 10 million images a month the API costs roughly $6,000 or more while one or two GPUs might cover it, so self-hosting wins if the model is good enough. Vision-language models are the most flexible and the most expensive per image; use them where the question changes or where a person would otherwise look, not on every video frame.

Video: sample frames, do not analyse all of them

Video multiplies everything. One camera at 30 frames a second produces over 2.5 million frames a day. Almost no use case needs all of them. Sample one or two frames a second, run a cheap motion or change detector first, and send only the interesting segments to the expensive model. In the AI body-camera operations platform we built for a US security services company, uploads are routed to specialist jobs for video stitching, audio extraction, frame and sprite-map generation and transcription, and Google Vertex AI reviews each incident rather than every frame. An hour of patrol footage becomes reviewable in seconds through the stitched video, frames and sprite maps.

Edge hardware

Running models on a device at the camera removes upload bandwidth and cloud GPU cost and keeps working offline, but adds hardware, enclosures, power and fleet management. NVIDIA launched its Jetson Orin Nano Super developer kit at $249 in December 2024 (Phoronix report of the launch); production modules, carrier boards, rugged enclosures, cameras and installation cost more per site. Budget engineering time for the fleet too: provisioning, certificate rotation, remote updates and health monitoring. On the body-camera platform, device fleet management uses AWS IoT provisioning and certificate rotation, and that work is a project in itself.

Monthly running cost checklist

• Inference: API fees per image or minute, or GPU hours whether used or not.
• Storage: original images and video, extracted frames and results for your retention period.
• Bandwidth: camera uploads and video playback, often bigger than compute for video products.
• Human review: people confirming alerts or low-confidence results.
• Re-labelling and retraining when products, sites or cameras change.
• Support and maintenance: $1.5k to $8k per month on our retainer, depending on scope.

How to keep the cost down

1. Try a managed API or a vision-language model on 200 of your own images before planning a custom model.
2. Fix the camera and the lighting if you can. A better image is cheaper than a better model.
3. Keep a person in the loop for the first release; automate once the measured error rate is acceptable.
4. Sample video frames and filter with cheap checks before calling expensive models.
5. Move to a self-hosted model only when monthly API spend or latency justifies a GPU.

For what works by industry, read computer vision use cases. For the wider machine learning picture, see machine learning app development cost. Use the app development cost calculator for the application around the model, and pricing for our rates. To staff it yourself, see hire AI and ML engineers.