All articles

Comparisons

Best face recognition APIs in 2026: 7 options ranked for photo search and face matching

Ranked for teams building face search or face matching into a product, with the source for every number.

By SightRadar EngineeringUpdated 9 min read Markdown

TL;DR

For face search and face matching inside a product, we rank SightRadar first on price per photo ($0.00062 in a batch), self-serve access and default rate limits. AWS Rekognition is the better choice if you are on AWS, run mostly 1:1 compares, or index more than about 37 million images a month. Azure Face suits approved Microsoft enterprise customers, and CompreFace is the pick when data must stay on your own servers.

Info: We build SightRadar, and it is first in this list. That makes this a vendor's ranking, so the criteria are stated up front, every number has a source, and each entry says where another option is the better choice.

Most lists of face recognition APIs rank on brand size, or on features copied from each vendor's docs. Neither tells you what happens when you build with one. Five plainer things decide a project:

  1. Can you sign up and call it today? Some APIs need an approval form before they will identify anyone.
  2. What does a real workload cost? Per image, per transaction and per month are three different bills.
  3. What are the default rate limits? A backfill of 200,000 photos at 5 requests a second takes eleven hours.
  4. How much code do you write? Especially if you already have AWS Rekognition code.
  5. Does it do recognition at all? Detection (where is a face) is not recognition (whose face it is).

This ranking is for one job: building face search or face matching into a product. That covers finding photos with a selfie, duplicate-account checks, and matching a selfie to a stored face. For on-device AR tracking, document KYC or data kept on your own servers, a different option wins, and each section says which.

The ranking at a glance

#APIType1:N searchSign up todayStarting price
1SightRadarCloud APIYesYes$0.00062 / photo (batch)
2Amazon RekognitionCloud APIYesYes, with an AWS account$0.00100 / image
3Azure FaceCloud APIYesApproval form$1 / 1,000 transactions
4Luxand.cloudCloud APIYesYes$9 / month for 1,500 requests
5Face++Cloud APIYesYesConcurrency packages
6KairosCloud APIYesYes100 free calls, then plans
7CompreFaceSelf-hostedYesNot applicableFree, plus your own servers
US list prices. Competitor pages read 2026-10-31; SightRadar rates from /pricing.

Google Cloud Vision is not ranked because it detects faces but does not recognise them. There is more on that near the end.

1. SightRadar: best for photo search and for teams leaving Rekognition

SightRadar is a face recognition API with the same collection model as Rekognition: create a collection, index photos, search with a selfie, compare two faces, detect faces. It does faces only.

  • Price per photo, not per face. $0.00062 a photo for batch indexing, $0.00093 real-time, $0.00062 for a search by stored face ID, and $6 per million stored faces a month. A photo with eight faces costs the same as a photo with one. See pricing.
  • Self-serve. Sign up, create a key, call it. There is no approval form and no cloud account to set up first.
  • Usable default limits. 50 requests a second for each key. A batch job takes up to 1,000 photo URLs in one call and answers by webhook, so bulk indexing is one request for every 1,000 photos.
  • Rekognition-shaped responses. In Python, a shim package keeps your boto3 method names, so the change is the import and the constructor. The migration guide has the before and after.
  • A dashboard for the whole thing. Collections, batch jobs, usage and billing are screens, not API calls, and a playground runs every operation in the browser.
  • Prepaid credits. No subscription and no minimum. When the balance runs out the API returns 402, and every billable response carries an X-Credits-Remaining header.

What it costs on a real workload. An event photo business runs 40 events a month at 2,500 photos each. Guests make 2,400 selfie searches, and 300,000 faces stay stored.

SightRadarAmazon Rekognition
100,000 photos indexed and 2,400 selfie searches$64.23$102.40
300,000 faces stored for a month$1.80$3.00
Total$66.03$105.40
Same workload at list prices. AWS figures are us-east-1, volume tiers included.

You can price your own volumes on four vendors at once with the cost calculator.

Where it loses:

  • Faces only. No labels, OCR, moderation, celebrity recognition, video or liveness.
  • 1:1 compares cost more than on Rekognition. A compare is billed as two photos. AWS bills a CompareFaces call as one image.
  • Very high volume. With Rekognition's volume tiers applied, its total bill drops below the batch rate above about 37 million images a month, and below the real-time rate above about 1.5 million.
  • The shim is Python only. Other languages call the REST API.
  • Accuracy is published for event photos. The benchmarks page gives the method and the limits. For KYC or access control, test on your own data first.

Pick it if you are building photo search, deduplication or selfie matching, or you have Rekognition face code you want to keep.

2. Amazon Rekognition: best on AWS and at very large volume

Rekognition is the default for a reason. It is mature, well documented, and its face collections hold very large galleries. It also does labels, text, moderation and video, which nothing else in this list does in one API.

Pricing (AWS pricing page, us-east-1): $0.00100 an image for the first million a month, then lower tiers down to $0.0004. Face storage is $10 per million faces a month. Some regions cost more.

What to watch:

  • Default quotas. IndexFaces, SearchFaces and SearchFacesByImage default to 5 transactions a second for each account in most regions, and 50 only in us-east-1, us-west-2 and eu-west-1 (service quotas). You can ask for more, but a first backfill often runs into this.
  • Input formats. JPEG and PNG only, 5 MB as raw bytes. iPhone HEIC uploads need converting first.
  • Console. AWS's guide says the Rekognition console "only provides demos" of features other than Custom Labels and Custom Moderation (AWS docs), so you list and delete face collections through the API or CLI.
  • You need an AWS account with IAM set up, which is real work if the rest of your stack lives elsewhere.

Pick it if you are on AWS already, you need labels or video as well as faces, your workload is mostly 1:1 compares, or you index more than about 37 million images a month. Our side-by-side is on SightRadar vs AWS Rekognition.

3. Azure Face: best for approved enterprise customers

Microsoft's Face API has detection, verification (1:1), identification (1:N) and find-similar. On features it is the closest match to Rekognition.

The catch is access. Identification and verification are Limited Access features. You fill in a registration form, and Microsoft decides. Detection works without the form.

Pricing (Azure Face pricing): $1 per 1,000 transactions for the first million, falling to $0.40 at high volume. Azure bills per transaction, not per photo. A selfie search is usually a Detect call plus an Identify call, so two transactions, and indexing a photo with three faces can be four. Count your real calls before comparing its price with a per-photo one.

Pick it if you are a Microsoft enterprise customer, your use case is approved, and your compliance team wants Azure. See SightRadar vs Azure Face.

4. Luxand.cloud: best for small, steady workloads on a fixed bill

Luxand has a hosted API for recognition, verification and detection, with extras such as age and gender estimates. It is quick to start with and has examples in many languages.

PlanPrice a monthRequestsRate limitStored faces
Start$91,50020 a minute500
Basic$3910,000100 a minute2,500
Pro$99200,000150 a minute6,000
Business$249500,000250 a minute18,000
Luxand.cloud plans, from luxand.cloud/pricing on 2026-10-31.

For photo search, the storage cap matters more than the request count. One 3,000-photo wedding produces about 9,000 face entries, which is more than the Pro plan stores.

Pick it if you have something like an office attendance app with a few hundred people and you want the same bill every month.

5. Face++: best for a wide attribute catalogue

Face++ (Megvii) is one of the oldest face platforms. Its catalogue is the widest here: detection, compare, search, dense landmarks, attributes, beauty scores and body analysis. The free tier has tight QPS limits, and paid throughput is sold as concurrency packages that you size in advance.

Note: Megvii has been on the US Entity List since 2019. Many US and EU procurement teams will ask about that, so raise it with yours early.

Pick it if you need its attribute and landmark range and your compliance team is comfortable with the vendor. See SightRadar vs Face++.

6. Kairos: best when the real job is identity verification

Kairos began as a face recognition API and now focuses on identity verification: document checks, selfie-to-ID matching and fraud signals. Its site offers 100 free API calls at sign-up.

Pick it if you are building KYC onboarding with documents. For photo search it is not the natural fit.

7. CompreFace: best self-hosted option

CompreFace by Exadel is open source under Apache 2.0. It runs with Docker Compose and gives you a REST API for recognition, verification and detection, plus a web UI for managing subjects.

What free still costs you:

  • A server to run it on, with a GPU if you need speed at volume
  • Upgrades, backups, monitoring and security patches
  • Scaling the vector search yourself as collections grow

Pick it if face data must stay on your own hardware and you have someone to run it. The trade-offs are in SightRadar vs self-hosted.

Not ranked: Google Cloud Vision and on-device SDKs

Google Cloud Vision appears in most lists, but for faces it does detection only. Google's face detection docs say "specific individual Facial Recognition is not supported". You get bounding boxes, landmarks and emotion likelihoods, with no collections and no matching. It is a good choice for blurring or cropping faces.

On-device SDKs are a different product. If you need live camera tracking for AR effects or virtual try-on, a cloud API is the wrong shape, because every frame would be a network round trip. None of the seven APIs above is built for that.

Which one should you pick?

Your jobPick
Photo search, selfie matching or deduplication in a productSightRadar, or Rekognition on AWS above about 37 million images a month
Mostly 1:1 comparesAmazon Rekognition
Faces plus labels, OCR or video in one APIAmazon Rekognition
Microsoft enterprise with approvalAzure Face
A few hundred people on a fixed monthly billLuxand.cloud
KYC with documentsKairos, or a dedicated identity verification vendor
Data must stay on your serversCompreFace
Only need to find faces, not identify themGoogle Cloud Vision

Two things to do before you commit

  • Test on your own photos. Public benchmark scores say little about your lighting, cameras and crowd sizes. Take a few hundred real samples with known answers and run them through your top two options. The threshold guide shows how to read the results.
  • Plan for re-indexing. Face vectors and thresholds do not transfer between engines. If you ever switch, you index again from the original photos and re-tune the threshold, so keep those photos.

SightRadar batch indexing is 38% below AWS Rekognition's first US tier, with self-serve keys.

See the face recognition API

Frequently asked questions

What is the best face recognition API in 2026?

It depends on the job. For photo search and face matching in a product, we rank SightRadar first on price per photo, self-serve access and default rate limits. AWS Rekognition is the better choice on AWS at very large volume or for 1:1 compares, and Azure Face for approved Microsoft enterprise customers.

Which face recognition API is cheapest?

At list prices under a million images a month, SightRadar's batch rate of $0.00062 a photo is 38% below Rekognition's first US tier of $0.00100. For 1:1 compares Rekognition is cheaper, because it bills one image for each compare. Self-hosting CompreFace has no per-call price, but you pay for servers and the time to run them.

What is the difference between face detection and face recognition?

Detection finds where faces are in an image. Recognition says whose face it is, by matching it against stored faces. Google Cloud Vision does detection only. The seven ranked options all do recognition.

Can I switch from AWS Rekognition without rewriting my code?

In Python, SightRadar's shim keeps the boto3 method names, arguments and response shapes, so the change is the import and the constructor. In other languages you call the REST API. With any other vendor it is a rewrite. In every case you index your photos again, because face vectors do not transfer.

Keep reading