AWS Rekognition cost: per image and per minute of video analyzed

Quick answer

Rekognition bills per image processed for image APIs (label detection, face analysis, moderation, text) and per minute of video for video analysis, at rates that vary by feature, with tiered discounts at volume and a free tier. Cost scales with how many images and minutes you analyze, so processing only what you need, sampling video frames rather than every frame, and caching results are the levers.

Amazon Rekognition analyzes images and video for labels, faces, text, and unsafe content. Its pricing is straightforward per unit analyzed, so the cost is entirely a function of how much you send it, and the common overspend is analyzing more than the use case needs.

Per image and per video minute

Analysis	Bills for
Image APIs (labels, faces, moderation, text)	Per image processed, by feature
Video analysis	Per minute of video, by feature
Face storage (collections)	Per face metadata stored per month

Each image API call bills per image, and each feature (labels, face detection, moderation) is priced separately, so calling several features on one image bills for each. Video is billed per minute analyzed. Rates fall in tiers at high volume, and a free tier covers small usage.

What drives the bill

High-volume image processing and, especially, per-frame video analysis drive cost. Analyzing every frame of a video is many times more expensive than sampling frames at intervals, which for most use cases (detecting when something appears) is sufficient. Calling more features than the use case needs also multiplies per-image cost.

Controlling Rekognition cost

Analyze only the images and features the use case requires, sample video frames rather than processing every frame where the detection allows, cache results so you do not re-analyze the same content, and pre-filter (skip obviously irrelevant images) before calling the API. As with any per-unit AI service, the entire optimization is sending less, the same discipline as the token levers in Bedrock.

FAQ

How is AWS Rekognition priced?

Per image processed for image APIs (label detection, face analysis, moderation, text), each feature priced separately, and per minute of video for video analysis, with tiered discounts at volume and a free tier. Face collections add a per-face-metadata-stored monthly charge. Cost scales with how much you analyze.

Why is Rekognition video analysis expensive?

Because it bills per minute of video, and analyzing every frame multiplies cost. Sampling frames at intervals, rather than processing every frame, is many times cheaper and sufficient for most use cases like detecting when something appears. Per-frame analysis is the common video-cost driver.

How do I reduce Rekognition cost?

Analyze only the images and features the use case requires (each feature bills separately), sample video frames rather than processing every frame where detection allows, cache results to avoid re-analyzing the same content, and pre-filter obviously irrelevant images before calling the API.

Does calling multiple Rekognition features cost more?

Yes. Each image feature, label detection, face analysis, moderation, text detection, is priced separately, so calling several on one image bills for each. Call only the features the use case needs rather than running all of them by default, to avoid multiplying per-image cost.

Does Rekognition have a free tier?

Yes, a monthly free allowance of images and video minutes that covers small usage and testing. Beyond it you pay per image and per video minute, with rates falling in tiers at high volume. The free tier makes low-volume experimentation essentially free.

Does C3X estimate Rekognition cost?

Rekognition cost is driven by images and video minutes analyzed, which are usage inputs. C3X prices the surrounding infrastructure, and you model your image and video analysis volume to estimate the per-unit AI charges.

What to do next

Estimate the infrastructure around your vision workload before you build it. C3X reads your Terraform and prices your resources against a live catalog. Start with the quickstart.