Face Recognition APIs and Open-Source Libraries in 2026: Prices, Licences and Traps

Two different jobs#

• Gallery matching: you enrol faces (users, employees, your photo library) and compare new photos against them, 1:1 (is this the same person?) or 1:N (who in my gallery is this?). Rekognition, Azure Face, Face++, Luxand and the open-source libraries do this. It costs fractions of a cent per call.
• Web face search: you send a face and get back public web pages where similar faces appear. Only a few services sell this as an API, it costs hundreds of times more per call, and it carries the legal exposure covered in our guide to face recognition laws. Our face search engine comparison covers the consumer side.

Price per 1,000 searches#

Cloud APIs#

Amazon Rekognition: cheapest self-serve option#

Amazon Rekognition bills per image per API: 1,000 SearchFacesByImage or CompareFaces calls cost $1.00 on the first tier, falling to $0.0004 per image above 35 million a month (pricing). Building the gallery costs the same per face indexed, plus $0.00001 per face per month for storage, so 100,000 enrolled faces cost about $1 a month to keep. Count calls, not searches: running DetectFaces and then SearchFacesByImage on one photo is two billed images. The AWS Service Terms carry a moratorium on police use of face comparison in criminal investigations (missing persons excepted), forbid sustained surveillance of a person without independent review, and add a biometric notice-and-consent addendum.

Azure AI Face: strong features, gated access#

Azure AI Face lists $1.00 per 1,000 transactions on its first Standard tier, down to $0.40 at 100 million or more (pricing), plus liveness at $15 per 1,000. Every call is a transaction and Microsoft's docs say detection comes first, so identifying a new photo takes at least two: about $2.00 per 1,000 searches. The bigger issue is access. Identification and verification are Limited Access: Microsoft decides at its own discretion, for customers it manages directly and for registered use cases, and not on the free tier. Emotion and gender detection are retired, and age, smile, facial hair, hair and makeup are restricted. Detection alone is open to anyone.

Face++: cheap, but check your compliance team first#

Face++ charges $0.002 per Compare or Search call pay-as-you-go, or $1,000 a month for one guaranteed query per second with unlimited calls, which only beats per-call pricing above 500,000 calls a month (pricing). Failed calls are billed too. Its maker, Megvii, has been on the US Commerce Department's Entity List since October 9, 2019, which matters for US companies and for many procurement and export-control policies.

Luxand.cloud: flat monthly plans#

Luxand.cloud sells monthly plans with request quotas: $9 for 1,500 requests ($6.00 per 1,000), $99 for 200,000 ($0.50 per 1,000) and up (pricing). Each plan also caps how many faces you can store, from 500 to 54,000, which is the limit most apps hit first. Its own pages disagree on the entry offer (a free 500 requests, a $5 plan and a $9 plan all appear), so confirm in the dashboard.

Open source#

DeepFace: easiest start, read the model licence#

DeepFace wraps 11 recognition models and 19 face detectors behind a few Python calls (verify, find, represent, analyze), and is MIT-licensed (repo). Its README warns that the licences of the models and detectors it wraps are inherited, and one of them, Buffalo_L, comes from InsightFace. Pick a model whose own licence allows your use. A hosted version, deepface.dev, charges about $2 per 1,000 images.

InsightFace: MIT code, non-commercial models#

InsightFace is the source of ArcFace, SCRFD and RetinaFace, and its code is MIT. Its pretrained model packs (buffalo_l, antelopev2 and others) are for non-commercial research only, whether they download automatically or you fetch them yourself. Commercial use needs a model licence whose price is not published. This is the trap that catches most teams: an MIT badge on the code says nothing about the weights.

CompreFace: self-hosted, but stalled#

CompreFace gives you a face recognition server with a REST API and a web UI in Docker, under Apache 2.0 (repo). Its default build uses FaceNet; custom builds based on InsightFace carry InsightFace's model terms. Its last release was v1.2.0 on August 22, 2023, so treat it as a starting point you will maintain yourself.

Web face search APIs#

Two consumer engines sell API access. FaceCheck.ID's API charges 3 credits per search at $0.10 per credit, about $300 per 1,000 searches, payable in crypto only, with a test mode that searches a small sample and is not meant for real results. Lenso.ai's Developer plan is $2,800 a month for 5,000 requests, about $0.56 each. Both search faces collected from the public web. Building a product on that is legal ground worth reading up on first: the EU AI Act bans building such databases by scraping, and Illinois' BIPA lets individuals sue over face data collected without written consent.

What to check before you ship#

• Model licence, not just code licence. InsightFace's weights are non-commercial, and DeepFace and CompreFace can pull them in.
• Billable calls per user action. Detect plus search is two calls on Rekognition and Azure; Face++ bills failures.
• Access approval. Azure identification and verification need Limited Access; budget weeks, not minutes.
• Consent and retention. In Illinois you need written consent and a published retention schedule before creating a face template. AWS and Microsoft both put notice and consent on you in their terms.
• Vendor status. Megvii is on the US Entity List; CompreFace is unmaintained.
• Laws by region. See our face recognition laws guide for the EU AI Act, GDPR fines, BIPA and Texas CUBI.