AWS Rekognition Pricing: Cost Per Feature (2026)
Nov 4, 202615 min read
Amazon Rekognition charges per call, per minute, per stored face, per session or per hour a model runs, depending on which feature you use.
The bill follows how many times you call each feature, not how many pictures you have, which is why teams that trigger calls from a workflow tool such as Activepieces still need to watch how often each step fires.
Amazon Rekognition is Amazon's image and video analysis service for objects, text, faces and unsafe content.
The table below lists each pricing line by what Amazon charges for and how the charge behaves. You can see which lines grow with traffic and which grow just by sitting there.
Amazon Rekognition pricing in one table
| API group | Billing unit | Volume tiers | Free tier | Usage limits |
|---|---|---|---|---|
| Group 1 images | Per image analyzed | Price steps down at higher monthly volume | Introductory allowance for new accounts | Billed per call, so two features on one image is two charges |
| Group 2 images | Per image analyzed | Price steps down at higher monthly volume | Introductory allowance for new accounts | Same per-call rule, at a different rate from Group 1 |
| Stored video | Per minute of footage | Price steps down at higher monthly volume | Introductory allowance for new accounts | Length of footage, not file count, sets the cost |
| Face storage | Per face stored, per month | None to plan around | Introductory allowance for new accounts | Charged every month, even if you never search |
| Custom Labels | Per training hour and per inference hour | None to plan around | Introductory allowance for new accounts | A running model bills even when idle |
| Liveness | Per verification session | Price steps down at higher monthly volume | None to plan around | Each retry is its own session |
You pay for actions, not assets. If you chain Rekognition into a workflow, every step that calls it is a billable call, so a flow that checks one photo three ways costs three times what you expected.
The free allowance is an introductory offer for new accounts, not a permanent discount. Treat it as a testing budget and price your real workload without it.
If you chain Rekognition into a workflow, every step that calls it is a billable call, so a flow that checks one photo three ways costs three times what you expected.
If you are running this arithmetic for your own team, see what the same workload costs on Activepieces.
Rekognition costs across volume tiers
Without the free tier, one Group 2 call per image costs $10 a month at 10,000 images, $1,000 at 1 million and $8,200 at 10 million, making large-scale image processing financially demanding. The table shows what happens when each image gets more than one call.
| Workload | 10,000 images | 1M images | 10M images |
|---|---|---|---|
| Moderation only | 10,000 calls × $0.001 = $10 | 1M × $0.001 = $1,000 | 10M calls at tiered rates = $8,200 |
| Moderation + labels + text | 30,000 × $0.001 = $30 | 3M × $0.001 = $3,000 | 30M calls at tiered rates: above $8,200, below $30,000 |
| Face search | 10,000 × $0.001 = $10 | 1M × $0.001 = $1,000 | 10M calls at tiered rates = $8,200 |
Run moderation, labels and text on every image and the call count triples. The 10,000-image bill becomes $30 and the 1 million bill $3,000, which means recognition costs are now significantly more expensive.
The search call costs the same as moderation, so the table shows the same figures. Face search also needs a collection, which is the stored set of faces you search against.
Amazon bills that stored face metadata monthly and separately. The line grows with every face you keep, even in a month when you run no searches.
Rekognition billing beyond per-image pricing
Stored faces are one of several charges that arrive without any new image to analyze. The others are each API call, Custom Labels hours, video minutes, storage, transfer and human review.
Billing per API call, not per image
Rekognition charges per call, and each type of analysis is its own call. Checking one photo for unsafe content, reading its text and detecting its labels is three billable operations on a single file.
Rekognition Custom Labels inference hour costs
Custom Labels lets you train a model to recognize your own products or defects. Amazon bills training by the hour, and bills inference too, meaning the hours a trained model is switched on. The meter runs whether you send it a thousand images or none.
Start a model for a demo and forget it over a long weekend, and it bills for the whole weekend.
Storage, transfer and human review around the call
Amazon bills video by the minute of footage analyzed, so a long recording costs more than a short one regardless of how little happens in it.
Your images also have to live somewhere, usually Amazon S3, which charges for storage and for data moved out. Any glue code that moves files between steps adds its own charges.
If you add Amazon Augmented AI, where a person reviews uncertain results, Amazon bills each review too. That cost rises with every borderline image, so a loose confidence threshold sends more work to reviewers than you planned.
Rekognition service limits and quotas that stop projects
Three limits stop most projects: how big an input can be, how many requests per second you may send, and how many faces a collection can hold.
Image and video input limits
Sending the image bytes in the request and pointing Rekognition at a file in S3 (Amazon's file storage) have different caps, and the S3 route allows larger files.
Either you resize before sending, or you store first and reference the file. Resizing is cheaper, but shrinking too far can hurt face detection on small faces in group photos.
Stored video has its own ceilings on file size and length. A long recording has to be cut into integrations before you submit it, and that splitting is work someone must build.
Default transactions-per-second throttling
Each API has a default ceiling on requests per second, and it differs by API and by region. Exceed it and Rekognition throttles you with errors instead of slowing down politely.
When a nightly job fires every call at once, it hits this wall. Without retry logic and a pacing step, it loses work silently.
Collection and face limits
A collection is the searchable store of faces you build. It has a ceiling on how many face records it holds, and a single image has a ceiling on how many faces get processed.
If you expect to outgrow one collection, plan to split by customer, site or year from the start, because merging them later is rework.
Most request-rate quotas can be raised through the AWS Service Quotas console or a support request. Hard limits on file size and format can't, so design around those.
Rekognition versus cloud and open source options
The three managed services all meter your usage. Open source on a GPU swaps the meter for a flat bill of about $384 a month, which you pay whether the machine is busy or idle.
| Price per 1,000 images (labels, text, moderation) | Free allowance | Faces | Custom models | Ops | |
|---|---|---|---|---|---|
| Rekognition | Billed per image, with a separate rate for each API group | Introductory tier for new accounts | Detect, compare and search | Custom Labels, billed for training and for hours running | AWS runs it |
| Google Vision | Billed per feature, per image | Monthly free units per feature | Detection, not identification | Trained through Google's wider ML platform | Google runs it |
| Azure AI Vision | Billed per call, by feature | Limited free tier | Identification is gated | Trained through Azure's vision tooling | Microsoft runs it |
| Open source on a GPU (about $384/month) | No per-image charge | Not applicable | Depends on the model and its license | Fine-tune anything you can host | You patch, scale and monitor |
I've left out the per-1,000 rates because they change often. Pull them from each vendor's own pricing page on the day you budget, and note that date in your estimate.
Face recognition policy differences
Face features are where the vendors differ most, and the difference is about permission rather than price. Google's Vision service detects faces but doesn't name them. Azure puts identification behind an approval process, so it may turn you down.
Rekognition has face search directly, and the legal responsibility for using it sits with you. Check your local biometric rules before you design around any of them.
When self-hosting beats a managed API
Self-hosting wins when your volume is steady and high enough that the flat GPU bill undercuts the meter, or when you need a model nobody sells. It loses when traffic is spiky, because an idle GPU still bills.
Small teams without someone to carry the pager should pick a managed service. Heavy, predictable workloads with an engineer on hand can pick open source.
Worth checking against a plan that does not meter every step: one credit covers a whole run on Activepieces.
Rekognition cost savings through single calls
Send each image to Rekognition once, ask only the questions you need, and let nothing sit running while you aren't looking.
- Call only the API groups you need. If you only want moderation labels, don't also run face analysis and text detection on the same file, because Amazon bills each as its own call.
- Resize images before sending them. A smaller file rarely changes the answer for labels or moderation, and it cuts upload time and storage.
- Dedupe first. Compare a file's fingerprint to ones you've already checked, so a re-uploaded photo reuses the earlier result instead of costing a second call.
- Sample video frames instead of scanning every one. You pay per call, so fewer frames means a smaller bill.
- Delete face collections you no longer search, since stored faces keep costing money while idle.
- Stop custom-label models when nobody is using them. A running model is billed for the hours it's up, whether or not a single image arrives.
- Set an AWS Budgets alert, so a runaway loop shows up as an email on day one instead of a surprise at month end.
Rekognition workflow steps in sequence
The controls above are easier to keep when the order of operations is fixed in one place. The canvas below shows a short chain of steps running in sequence, each doing one job, with loops marking where state is carried forward.
Your version swaps the chat steps for a new-file trigger, one Rekognition call, and a branch.
Flagged items go to a human reviewer, and clean ones are saved with their result. Because the result is stored, no later step needs to call Rekognition again for the same file.
Activepieces meters the run, not the steps inside it: its published pricing page lists 1 credit per flow run, so a ten-step flow costs the same as the two-step shortcut it replaced.
Zapier's pricing page bills per task and Make's per module, so there every extra step, including the small, separate ones that make a flow reliable, raises the count. AWS still bills the single Rekognition call separately.
Rekognition spend estimation in one week
Spend one Monday on the steps below and you'll have a monthly estimate built on calls, not images. That's the only version that survives contact with your first invoice.
Rekognition cost estimate in five steps
Run the list below in order, because each step feeds the next. Skip the early ones and the later ones multiply a wrong number, so you'll be confidently off by a wide margin.
- Count what you process in a month: still images in one pile, minutes of video in another. Use real upload logs, not a guess from memory, because the two are billed on different meters and your totals decide which price list applies.
- For each image, list every Rekognition API you intend to call: faces, labels, text, moderation, whatever the workflow needs. Five checks on one photo are five billable calls, so this list is what turns a small image count into a large call count.
- Multiply each call count by the price for its API group and your volume tier, taken from AWS's pricing page. Add the results. Do video separately, per minute.
- Add the costs around the calls: storage for the files and stored results, and the human review of anything the service flags as uncertain. People forget these because they aren't on the Rekognition line, but they land on the same bill.
- Check Service Quotas in your AWS account for the limits on each API you plan to use. A quota that is too low means your job stalls mid-run, which costs you time even when it costs no money.
Then set a budget alert in AWS Budgets at a level just above your estimate. An overrun then reaches you as an email and not as a surprise at month end.
Last, re-verify every price on the AWS Rekognition pricing page before you commit. Prices change, and an estimate built on a remembered number is only as good as your memory.
What Activepieces does about this
The cost problem in this article is that the bill follows calls, and calls multiply quietly inside workflows. Activepieces can't lower AWS's rate, and the Rekognition call is still billed by AWS.
What it does is let you fix the number of calls per image in one visible place, so the estimate you built in the five steps above is the number the workflow actually produces.
A flow has one trigger, such as a new file landing in a bucket. It then makes one Rekognition request, branches on the result, and stores the output.
Because the stored result sits in the flow's own path, nothing downstream has a reason to ask Rekognition about the same file again.
A dedupe check before the call, using a file hash against results you already hold, is one more step in the same flow. If the integration library lacks the exact Rekognition action you need, the HTTP step can call the API directly.
The pricing is the part you can check. Activepieces' published pricing page lists 1 credit per flow run, so adding the dedupe check, the branch, or the reviewer notification doesn't enlarge the automation bill.
Per-task and per-module tools charge for each of those steps, and Activepieces doesn't. That matters here because the steps that make a Rekognition pipeline cheap are the small, separate ones.
The core is open source under the MIT licence, so you can self-host it next to your AWS account and keep images, results and credentials on infrastructure you control.
Self-hosting brings the same trade-off as the GPU option earlier: you carry the patching and the pager. For a team without that capacity, the hosted version trades control for no operations work.
Either way, the estimate stays what it was: calls per month, priced at AWS's rates, with the workflow held to the call count you planned.
Frequently asked questions
Is Amazon Rekognition free to use?
Not indefinitely, but AWS has offered a free tier for new accounts that covers a limited monthly volume of image analysis for a set period. Once it ends, or once you pass its limits, you pay per call.
Treat the free tier as a way to test your design, not as a budget line, and check the current terms on the AWS pricing page before you rely on it.
Pricing basis
Does Rekognition charge per image or per API call?
Per API call, which in practice means per image per call. Send one photo through moderation, face detection and text detection, and you've made three billable calls, not one.
That's why this article keeps telling you to estimate by calls per month. Counting images alone will undercount the bill by however many analyses each image goes through.
How much does it cost to moderate 100,000 images?
Multiply 100,000 by the current per-image moderation rate for your region, then add every other call each image triggers. I won't print a total here, because a remembered number goes stale and a wrong one goes into your budget.
Do the multiplication with the live price, and add storage and any human review of flagged images on top.
Regions and discounts
Does Rekognition pricing differ by region?
Yes. AWS sets prices region by region, so the same call can cost more in one than another.
If your images already live in a particular region, price it there first. Moving data to a cheaper region can cost you more in transfer and delay than you save on the call.
Can I get a discount for committed usage?
AWS has historically tiered image pricing so that the rate drops as monthly volume climbs, and large customers can negotiate private pricing with their account team.
If your volume is high and steady, ask before you build the estimate. If you're small, plan on the standard rates and don't count on a discount.
Related reading
References
Running the numbers
See what the same workload costs here.
Free forever plan, and every paid plan self-hosts at no extra cost.
See pricing Talk to sales
