Click every traffic light: what CAPTCHA and reCAPTCHA really cost

Where the seconds you spend on reCAPTCHA have gone since 2007, what one university study estimates they add up to, why bots now pass the test more reliably than people, and when a contact form needs a CAPTCHA at all.

Everyone has done it: ticked "I’m not a robot", clicked every square with a traffic light in it, and then done it again because a corner of the pole apparently counted. A CAPTCHA is meant to be a small toll for using a website. This guide follows where that toll went: how reCAPTCHA turned the seconds people spent on it into digitised books, house numbers for Google Maps and labelled images, what a 2023 UC Irvine study estimates those seconds add up to, and why later research finds that bots now pass the test more reliably than people do. It also covers what a CAPTCHA costs the website that runs it, in lost enquiries, accessibility and privacy law, what changed when Google moved reCAPTCHA into Google Cloud, and when a contact form needs one at all. The figures come from the papers and vendor pages listed at the end, read on 3 November 2026.

CAPTCHA and its alternatives, side by side

OptionRun byWhat the visitor doesWhat it checksPrice, November 2026
reCAPTCHA v2GoogleTicks a box; picks images if in doubtThe interaction, cookies and browser history10,000 checks a month free, then billing
reCAPTCHA v3GoogleNothing; the site gets a score from 0.0 to 1.0Interactions across the site10,000 checks a month free, then billing
Google QR code challengeGoogleScans a QR code with a phoneThe phone: Google Play services, or iOS 16 with an app or App ClipPreview, on request
Cloudflare TurnstileCloudflareUsually nothing; sometimes a checkboxProof-of-work and browser checksFree, up to 20 widgets
hCaptchaIntuition MachinesMay get a challenge; passive mode on paid plansRisk signals, then the challenge answerFree; Pro from $99 a month, billed yearly
Friendly CaptchaFriendly Captcha, GermanyNothing; the device solves a puzzleProof-of-work, with no cookiesFree for small non-commercial sites; from €9 a month
Private Access TokensThe device maker and a token issuerNothingThat the device is genuineBuilt into iOS 16 and macOS 13
Honeypot fieldYour own siteNothingWhether a hidden field was filled inFree
Rate limitingYour own serverNothing, unless they flood the formHow many requests come from one sourceFree

From each vendor’s own pages, read on 3 November 2026. reCAPTCHA’s free allowance is per Google Cloud organisation; from 10,001 to 100,000 checks a month it costs a flat $8, then $1 per 1,000. Friendly Captcha’s free plan covers one domain and 1,000 requests a month. Honeypots and rate limiting are techniques, not products.

A poll, two universities and the first bots

In November 1999 Slashdot ran an online poll asking which was the best graduate school in computer science. Students at Carnegie Mellon wrote programs that voted for CMU thousands of times. The next day students at MIT wrote their own, and the poll turned into a contest between bots that MIT won by 21,156 votes to 21,032. Every other school finished below 1,000. The story opens the 2003 paper that formally introduced the CAPTCHA, by Luis von Ahn, Manuel Blum and Nicholas Hopper of Carnegie Mellon and John Langford of IBM, who had first proposed the idea in 2000.

The name stands for Completely Automated Public Turing test to tell Computers and Humans Apart: a test a computer can set and mark, which most people pass and the programs of the time could not. The authors described it as a win-win. Either the test holds and a site can tell people from programs, or someone breaks it and in doing so solves a hard problem in artificial intelligence. Yahoo! was already using one of their CAPTCHAs to stop bots from opening thousands of free email accounts a minute.

reCAPTCHA: ten seconds of reading, put to work

By 2008 people were solving more than 100 million CAPTCHAs a day, which Carnegie Mellon put at hundreds of thousands of hours of human effort, all of it thrown away once the box was ticked. Von Ahn’s answer was reCAPTCHA, built from images of words that optical character recognition had failed to read in old texts. Each time someone typed one, a word got transcribed. In its first year it deciphered more than 440 million words, and by August 2008 more than 4 million a day, for books in the Internet Archive and the archive of The New York Times.

The results reached better than 99% accuracy, the standard expected of professional human transcription, on faded pages where OCR could miss as many as one word in five. On 16 September 2009 Google announced it had acquired reCAPTCHA, under the headline "Teaching computers to read". From then on, all that reading went to work for the company behind Google Books.

House numbers, then traffic lights

In March 2012 people began seeing house numbers next to the usual distorted words. Google confirmed to TechCrunch that the images came from Street View, and said it normally extracts details such as street names and traffic signs from that imagery to improve Google Maps with business addresses and locations. In April 2014, members of Google’s Street View and reCAPTCHA teams published a neural network that read house numbers from Street View with over 90% accuracy. Applied to reCAPTCHA’s own distorted text, it transcribed the hardest category at 99.8%.

The test had been beaten by the team that ran it. In December 2014 Google introduced the "No CAPTCHA reCAPTCHA" checkbox, saying that artificial intelligence could now solve even the most difficult distorted text at 99.8% accuracy, and began testing image challenges, such as picking the pictures of cats. The grids of buses, crosswalks and traffic lights followed. Whether those clicks train self-driving cars is the most repeated claim about reCAPTCHA, but none of the sources below confirms it: Google’s documented use of reCAPTCHA images is Maps, and the ETH Zurich researchers who broke reCAPTCHA v2 in 2024 write only that it is "suspected" to be useful for self-driving cars.

What the checkbox is really reading

The 2014 checkbox looked like a test of whether you can click a box. Google described it differently: an advanced risk analysis engine that weighs the user’s whole engagement with the CAPTCHA, before, during and after the click, and serves an image challenge only when it is unsure. reCAPTCHA v3, launched in October 2018, dropped the challenge altogether. It returns a score from 0.0, very likely a bot, to 1.0, very likely a good interaction, with 0.5 as the default threshold, and Google recommends running it in the background of pages as well as on forms, because it "works best when it has the most context about interactions with your site".

The checkbox is not the test; the browser behind it is. In 2024 ETH Zurich researchers found evidence that reCAPTCHA v2 relies heavily on cookies and browser history to decide whether someone is human: adding a real user’s history to their bot’s browser cut the number of challenges it had to solve. Google’s own FAQ says reCAPTCHA sets a cookie, _GRECAPTCHA, which it calls necessary for its risk analysis. The same researchers point out the other side: some people simply do not have convincing enough browsing data, and they are the ones sent back to the image grid.

The bill: 819 million hours, by one estimate

In 2023 Andrew Searles, Renascence Tarafder Prapty and Gene Tsudik of UC Irvine published "Dazed & Confused", a 13-month study of more than 3,600 real users who met reCAPTCHA v2 on a live account and password recovery service at their university, and who did not know they were being studied. Their users rated the checkbox "good" on the System Usability Scale, at 77, and the image challenge only "OK", at 59; they also found the image challenge annoying. The paper then estimates the cost of reCAPTCHA worldwide, assuming 100 million challenges a day, the figure reCAPTCHA itself gave in 2010.

On that assumption, the study counts at least 512 billion reCAPTCHA sessions between versions 1 and 2, taking 819 million hours of human time, which it values at $6.1 billion in unpaid wages at the US federal minimum wage. reCAPTCHA v2 alone, at 408 KB a session without caching, moved 134 petabytes of data, which the authors put at about 7.5 million kWh of energy for network transmission only, and 7.5 to 18 million pounds of CO2, roughly 3,400 to 8,200 tonnes. They describe all of these as generous lower bounds.

The two remaining estimates are about value rather than cost, and they are the study’s most speculative. Priced at Google’s own data-labelling rates of $35 to $129 per 1,000 items, the image labels collected by reCAPTCHA v2 would be worth $8.75 to $32.3 billion each time the set was sold. Valuing each of the 329 billion reCAPTCHA v2 sessions at an average cookie lifetime value of about $2.70, the authors arrive at $888 billion in tracking cookies. Their conclusion is that reCAPTCHA v2 "has an immense cost and no security", and that it should be deprecated.

The machines got better at it than we did

The "no security" half comes from a second UC Irvine paper, presented at USENIX Security 2023 with researchers from ETH Zurich and Microsoft. 1,400 participants solved 14,000 unmodified CAPTCHAs of the kinds real websites use. People got 50 to 85% of them right, depending on the type. The automated solvers published in the research literature reached 85 to 100%, most of them above 96%, and were faster in every case except reCAPTCHA’s image grid, where people took 18 seconds against the bots’ 17.5; inside a realistic sign-up task, people took 22.

On those figures, the test now fails more humans than machines. In September 2024 Andreas Plesner, Tobias Vontobel and Roger Wattenhofer of ETH Zurich used off-the-shelf YOLO image models to solve 100% of reCAPTCHA v2 challenges, where earlier work had managed 68 to 71%, and found no significant difference between the number of challenges their bot and people had to solve before passing. Attackers have noticed: Google’s own threat intelligence group has documented CAPTCHA solving built into automated account farms, as our guide to how AI agents changed hacking describes.

Google’s answer: Cloud billing and a QR code

Google has stopped issuing classic reCAPTCHA keys. New ones were no longer allowed from the third quarter of 2024, existing keys were moved into Google Cloud projects automatically from the fourth quarter of 2025, and by the first quarter of 2026 API access was locked for keys without a project. The free allowance is now 10,000 assessments a month per organisation. From 10,001 to 100,000 a month the Premium tier costs a flat $8, and beyond that $1 per 1,000. A site that goes over the free allowance without enabling billing gets an odd result from the classic verification endpoint: Google’s migration guide says those requests "fail open", returning success with a score of 0.9.

On 22 April 2026 Google introduced Google Cloud Fraud Defense, which it calls "the next evolution of reCAPTCHA", built to tell apart bots, humans and AI agents, with reCAPTCHA as its core bot defence and no change for existing keys or prices. It added a new challenge that Google calls AI-resistant: a QR code the visitor scans with their phone. The phone then verifies in its own secure hardware, which on Android requires Google Play services version 25.41.30 or later and on iPhone requires iOS 16 or later with Google’s reCAPTCHA app or an App Clip. After 26 years, proving you are a person has gone from typing a word to owning the right phone.

What a CAPTCHA costs the site: enquiries, accessibility and the GDPR

Every challenge is an extra step at the exact moment someone has decided to contact you, sign up or pay. In the USENIX study, 30% of paid participants gave up before finishing, and those who met the CAPTCHAs inside a realistic account sign-up were 120% more likely to give up than those solving them on their own. It is the same arithmetic as a long checkout, where shoppers resent typing more than clicking, as our guide to why customers leave checkout shows: anything a visitor has to do that does not help them is a reason to stop.

Accessibility is the second cost. WCAG 2.2 success criterion 3.3.8 forbids a cognitive test such as a puzzle at any step of a login unless an alternative or an exception applies; recognising objects, the traffic-light grid, is excepted at level AA but not at AAA. The W3C’s note on the inaccessibility of CAPTCHA lists who the usual fallbacks fail: visual puzzles exclude blind users, audio ones exclude deaf and deaf-blind users, few audio CAPTCHAs can be replayed reliably, and they are a poor choice for people with cognitive disabilities. It also records that reCAPTCHA’s audio challenge was broken using Google’s own speech recognition. What the European Accessibility Act now requires is in our guide to WCAG 2.2 and the EAA.

In the EU, reading or storing anything on a visitor’s device needs their consent unless it is strictly necessary for the service they asked for, the rule explained in our guide to cookie banners and Consent Mode. Google calls the reCAPTCHA cookie necessary. France’s data protection authority, the CNIL, has taken a different view: in a decision of 16 March 2023 it fined the scooter-rental company Cityscoot €125,000 for several breaches, one of which was running reCAPTCHA without the consent the law requires, while the mechanism collected hardware and software details from users’ devices and shared them with Google.

One national decision is not a ban, but it sets out where the risk sits: with the website, not with Google. A site that loads reCAPTCHA before consent is relying on that device data being strictly necessary, and the CNIL did not accept that argument. Consent itself has to be freely given and as easy to refuse as to give, the same standard that a cookie banner with a buried reject button fails, as our guide to dark patterns explains. The simpler way out is a check that reads nothing from the device in the first place.

The alternatives, and what each one reads

Cloudflare Turnstile, launched in September 2022, runs small non-interactive checks in the background, proof-of-work and browser tests among them, and shows a checkbox only to some visitors. It is free on any website, up to 20 widgets, without routing traffic through Cloudflare, and Cloudflare says it never looks for cookies such as a login cookie. hCaptcha, run by Intuition Machines, lists a low-friction passive mode on its paid Pro plan, from $99 a month billed yearly, and a no-CAPTCHA mode on Enterprise; its free Basic plan lists neither. Friendly Captcha, a German company, has the visitor’s device solve a cryptographic puzzle while they fill in the form and says it stores no cookies, trackers or personal data; its free plan is limited to small non-commercial sites, and EU-only data centres come with its Advanced plan.

Private Access Tokens, built into iOS 16 and macOS 13, let the device maker vouch that a real device is asking without telling the website who the person is; Cloudflare expected them to remove nearly every CAPTCHA served to supported Apple devices. The W3C note recommends the same direction for everyone else: non-interactive checks first, honeypot fields that only bots fill in, spam filtering and limits on what new accounts can do, and a choice of challenge types where one is unavoidable.

Does your contact form need a CAPTCHA?

For an ordinary contact form, start with the checks a visitor never sees. A well-built form has a honeypot field hidden from people, rejects submissions sent faster than anyone could type them, limits how many messages one address can send in an hour, validates every field on the server and filters what gets through before it reaches an inbox. None of it asks the visitor for anything, none of it needs a cookie, and a contact form whose worst case is a few spam emails a week rarely needs more.

A real challenge earns its place where each automated attempt costs you money or exposes an account: sign-ups, logins targeted with stolen passwords, forms that send SMS messages, free trials, ticket sales. There, pick an invisible check that challenges only suspicious traffic, put it only on that step rather than on every page, and decide before launch how it will handle consent and visitors who cannot see or hear a puzzle. If you are not sure what your forms load and what protects them, a Technical Audit reviews the site’s security and the third-party code it depends on, and puts what to fix first at the top of the list.

Frequently asked questions

Sources and method

Every figure is from the papers and pages below, read on 3 November 2026. The cost estimates are the UC Irvine authors’ own, quoted as they give them; the only conversion is pounds of CO2 to tonnes. Prices are each vendor’s published list prices on that date.

Spam getting through your contact form?

We build forms that stop automated spam with checks your visitors never see, and add a challenge only on the steps where abuse actually costs you something.