Facial Recognition Software for Law Enforcement: What Agencies Actually Paid in 2026

Last updated: October 13, 2026

Eight US law enforcement agencies paid between $7,000 and $62,500 for Clearview AI facial recognition software in contracts recorded between August 2025 and September 2026. The median was $25,814. Bigger biometric systems ran higher, up to $1,077,949. This is a small sample from public records, not a market rate. It is also more than any vendor page on this subject will tell you.

What agencies paid for facial recognition software in 2026

A city police department pays roughly $13,000 to $62,000 a year for facial recognition software. That is the range in the records. It is narrower than most people expect. One product dominates this end of the market.

Source: Civic IQ contract records, council agendas, purchase orders and check registers, August 2025 to September 2026. Eight agencies. This is a small sample and skews toward small and mid-sized agencies.

Two things in that table matter more than the numbers.

Almost none of these went to competitive bid. Southaven and East Point both recorded sole-source buys. Amarillo was approving a fourth one-year renewal in a row. Denton County signed a 36-month term. Once a department has this software, the choice is whether to renew. It is no longer which vendor to pick.

Some of it does not look like a contract at all. The Fishers, Indiana figure is a line in an accounts payable check run, coded as professional services. If you want to know what your peers pay, the council agenda is often not where it is written.

Larger biometric systems cost far more

Some agencies buy facial recognition inside a full identification system. That is a different order of spending from a search subscription.

Source: Civic IQ contract records, 2025 to 2026.

The McDonough record is the one to study if you are budgeting. Facial recognition arrived bundled inside an analytics buy, not as its own line. That is increasingly how it is sold. It is also why a department can be using the technology without a council ever voting on facial recognition by name.

Fingerprint and livescan are a separate, cheaper category

Biometric identification is not one market. Fingerprint and livescan systems are booking-desk hardware. Agencies have run them for decades. Those came in at a median of $13,415 across eight agencies, from $3,102 to $46,244.

Source: Civic IQ contract records, 2025 to 2026. Eight fingerprint and livescan awards in total; five shown.

Keep these apart when budgeting. A livescan swap is a hardware refresh on a known cycle. A facial recognition subscription is a repeating software cost with a policy question attached.

The vendors public agencies actually buy from

Clearview AI appears in more of these records than every other facial recognition vendor combined. That is the finding. It does not match the vendor lists that rank for this search term.

Source: Civic IQ contract records, August 2025 to September 2026. Vendors named as recorded in the awarding document.

Three names fill the vendor roundups for this term and appear in none of these priced awards: NEC NeoFace, Amazon Rekognition and Microsoft Azure Face API. That says something about this sample, not about those firms. NEC in particular has supplied large agencies for years. Large-agency contracts are less likely to show up in the council-agenda records this sample draws on.

For a buying team the practical read is simple. The vendor list you will actually choose from is shorter than the internet suggests. At city scale it is often one incumbent and a renewal date.

How facial recognition software works, and where it fails

Facial recognition turns a face into a set of numbers. It then compares that to others. There are two jobs here. Agencies buy them for different reasons.

• One-to-one verification. Does this face match this one record? Used at a checkpoint, or to unlock a phone. One comparison.
• One-to-many identification. Who is this? It searches a gallery. That gallery may be a booking database. In Clearview’s case it is images taken from the public internet. Millions of comparisons.

Police buy the second. The difference matters. The errors behave differently. A one-to-many search returns a candidate list. A candidate is not a match.

What NIST actually measured

NIST runs two evaluation programmes for this technology. Identity matching runs under FRTE. Image analysis runs under FATE. Both replaced the older FRVT name. Vendors submit their algorithms. NIST keeps publishing results. So treat any accuracy claim you cannot trace to NIST as marketing.

For a public agency the most useful document is still NISTIR 8280, the December 2019 report on demographic effects. It tested nearly 200 algorithms from about 100 developers. It used more than 18 million images of over 8 million people. The findings are specific, so they are quoted here rather than summed up.

• On higher quality photos, NIST found “false positive rates are highest in West and East African and East Asian people, and lowest in Eastern European individuals.” It called the effect “generally large, with a factor of 100 more false positives between countries.”
• On the images agencies actually use: “With domestic law enforcement images, the highest false positives are in American Indians, with elevated rates in African American and Asian populations.”
• On sex: “We found false positives to be higher in women than men.” NIST called this smaller than the effect due to race.
• On age: false positives rose in the elderly and in children. They were largest in the oldest adults and the youngest children.

One finding deserves real attention at procurement. NIST noted that several algorithms built in China returned low false positive rates on East Asian faces. Where an algorithm was built stands in for the make-up of its training data. So two products with the same headline accuracy can fail in different ways on your population.

Note the date. That report is from 2019 and algorithms have moved on since. It is still the most rigorous public measurement of the problem. No vendor has published anything close to it.

Why the accuracy numbers on this subject are unreliable

Search this term and you will be told facial recognition is “99.9% accurate.” Or that it beats “95% accuracy under optimal conditions.” Neither figure carries a benchmark, a dataset or a source. Both sit on pages ranking on page one today.

An accuracy figure means nothing without four things. Which algorithm. Which dataset. One-to-one or one-to-many. And at what threshold. A number without those is not a measurement. This article repeats none it cannot trace to NIST.

Who owns the vendor, and why a buyer should care

The company that dominates these records is 23 percent owned by the people who sued it. That is odd enough to be worth a buyer’s attention.

On March 20, 2025 a court approved a settlement in the class action over Clearview AI’s collection of face images. Class members did not get cash. They got a 23 percent equity stake in the company. The stake was valued at about $51.75 million. That rested on an agreed company value of $225 million as of January 2024.

The settlement also says what happens if no sale or listing follows. A court-appointed monitor can make the company pay 17 percent of its revenue over two years into a settlement fund.

Read that against the contracts earlier in this article. The revenue in question is agency subscription revenue. The money a police department pays for this software sits in the pool a monitor can draw on.

What it means at procurement

Nothing here says the product stops working. It does say three practical things.

The vendor’s ownership and finances are an open question in a way most suppliers’ are not. A five year commitment deserves the same solvency questions you would put to a small vendor. Ask what happens to your data and your case records if the company is sold. Get the answer in the contract.

It also explains the shape of the market here. Eight of eleven contracts went to one supplier, mostly without competition. The established names did not appear at all. One dominant seller and an unsettled owner leaves a buyer with less leverage, not more.

Is facial recognition software legal for police in 2026?

It depends on the state, and in some states on the city. There is no single answer. The position has also moved in the past year. What follows is where things stood in September 2026. It is stated narrowly, because this area changes fast.

The clearest recent change is in Virginia. It is worth reading closely, because it shows how these rules actually work.

Virginia, effective July 1, 2026

Code of Virginia section 15.2-1723.2 took effect on July 1, 2026. A local police agency may buy or use facial recognition only where doing so “is expressly authorized by statute.” section 23.1-815.1 applies the same rule to campus police.

The statute sets conditions on any use it does allow. The agency must keep the technology under its own exclusive control. The data must stay confidential. It may not be shared or resold. And it can be reached only by search warrant or administrative warrant. Commercial air service airports are excepted.

One detail matters for handling evidence. The version in force before July 1, 2026 said a match “shall not be included in an affidavit to establish probable cause.” The version now in effect drops that line. Agencies in Virginia should take their own legal advice on what that means. Do not assume the old rule still holds.

The wider pattern

Beyond Virginia, states fall into a few groups. Some require a warrant, probable cause or a court order before police may run a search. Several bar a match from being the sole basis for an arrest. Several require notice that the technology was used. And a number of cities ban city government use outright, starting with San Francisco in 2019.

Published counts of how many states sit in each group differ between trackers. They sort partial restrictions in different ways. So no count is quoted here. Before buying, read your own state code, then your city ordinance. Check whether any existing authorisation carries an end date. Virginia’s did.

The contract records show buying carrying on regardless. Lawton, Oklahoma funded its acquisition through a grant. Southaven recorded a sole-source purchase. Buying and policy are moving on separate tracks. In several of these records the software arrived before any published use policy did.

What “advanced” and “AI” facial recognition means in 2026

Mostly two things. Liveness detection, and matching that holds up on poor images. Neither is what most agency buyers actually need.

• Liveness detection and anti-spoofing. Checking that the face at the camera is a live person, not a photo, a mask or a replayed video. This matters for access control and identity checks. It matters much less to an investigator searching a still image against a gallery.
• Uncontrolled capture. Matching from off-angle, low-light or partial images. This is where the research effort now sits. In practice it is what separates products.
• Face analysis, not recognition. Estimating age, mood or expression. NIST tests this separately under FATE, because it is a different problem.

Facial emotion recognition sits in that last group. It is worth one sentence only, because it does not appear in these records at all. No agency in this sample bought it. Treat any pitch built on reading emotion as outside the evidence base, not just unproven.

How to evaluate a facial recognition bid

Most agencies in these records did not run a competitive process. If you do, these questions separate a serious bid from a brochure.

• Has the algorithm been submitted to NIST, and under which name? Vendors often submit under a developer name that differs from the product name. Ask for the exact name, so you can look the results up yourself.
• One-to-one or one-to-many, and against what gallery? A subscription that searches public internet images is a very different thing, in law and in practice, from one that searches your own booking photos.
• What demographic testing has been done on your population? NIST found errors vary with where an algorithm was trained. One headline accuracy figure does not answer this.
• What is the candidate list threshold, and who sets it? The threshold sets how many false matches an investigator sees. If the vendor controls it, you do not control your own false positive rate.
• What is the audit trail? Who searched, for what, under what case number, and kept for how long. In several states this is now the law, not just good practice.
• What happens at renewal? These contracts renew as a matter of routine. Amarillo was on its fourth one-year renewal in a row. Cap the price rises at the start.

And one question for the agency, not the vendor. Can your department say in writing, before it buys, what a match may and may not be used for? Several records here show the software bought first and the policy written later.

Frequently asked questions

How much does facial recognition software cost?

Eight police agencies paid between $7,000 and $62,500 for Clearview AI in records dated August 2025 to September 2026. The median was $25,814. Bigger biometric systems cost far more. One county airport system came to $1,077,949. Fingerprint and livescan gear is cheaper, at a median of $13,415.

What is the best facial recognition software for law enforcement?

The records do not support naming one. Clearview AI appears in 8 of 11 contracts here. But being common is not the same as being good. The fair way to compare is to look up each vendor’s algorithm in the NIST results. Then check how it does on your image types and your population.

Is facial recognition software illegal?

Not as a rule, but it is restricted in a growing number of states and banned for city use in some places. Virginia is the clearest recent case. Since July 1, 2026 a local police agency there may use it only where a statute says it can. Check your state code, then your city ordinance.

Can facial recognition be used to find someone?

Police use it to build leads, not to identify people outright. A one-to-many search returns a candidate list, and a candidate is not a match. Several states bar a facial recognition result from being the sole basis for an arrest. Some also require notice that it was used.

How accurate is facial recognition software?

There is no single number. Treat any vendor quoting one without a benchmark with care. NIST tested nearly 200 algorithms in 2019. It found false positive rates vary sharply by group. On police images, the highest were in American Indians. Rates were also raised in African American and Asian groups.

Do agencies competitively bid facial recognition software?

Usually not. In this sample the buys were recorded as sole source, as renewals, or as payments in a check register. Amarillo was approving a fourth one-year renewal in a row. Once the software is in use, the question becomes whether to renew it, not which vendor to choose.

Related reading: our analysis of licence plate reader alternatives for government agencies, what US cities actually bought in smart city technology, what districts paid for school security systems, and the Verkada government camera market.

On the plate-reading side of this market, see the Flock camera map and contract values.

Contract figures come from individual records surfaced by Civic IQ from council agendas, purchase orders and check registers, pulled September 13, 2026. Accuracy findings are quoted from NISTIR 8280 and are dated 2019. Statutory text is quoted from the Code of Virginia. No vendor-published accuracy claim is repeated anywhere in this article. Civic IQ tracks 100,000+ agencies, $14T+ in tracked spend, 1.5M+ documents per month, all 50 states.