reCAPTCHA vs hCaptcha vs Turnstile for an AI agent

reCAPTCHA, hCaptcha and Cloudflare Turnstile compared by what each asked of an AI agent's browser in August 2026: a score, a puzzle, 35 seconds or no click at all.

Davide Valli · Founder, Scalebrowser

Published Nov 4, 2026 · 6 min read

Short answer

reCAPTCHA, hCaptcha and Cloudflare Turnstile differ in what they ask of an agent: Turnstile in managed mode returned a token in 1.7 seconds with no interaction at all, reCAPTCHA v2 opened the same image puzzle for our agent's browser as for ordinary Chrome and the agent solved it, and hCaptcha took about 35 seconds per set of two image pages, 4 of 4 runs, measured in August 2026. None is harder in general: the site key and the difficulty its owner sets decide more than the vendor.
35 seconds, per hCaptcha set, beside an hourglass with sand below

A site that switches its challenge vendor changes what your agent has to do on the next run, from nothing at all to half a minute of image tasks. Comparisons of the three usually rank them by privacy or price for the site owner. We compared them by what each actually asked of an agent's browser on live pages with production keys between 4 and 14 August 2026.

What are reCAPTCHA, hCaptcha and Turnstile?

reCAPTCHA, hCaptcha and Cloudflare Turnstile are the three challenge services most sites embed to decide whether a visitor is a person, and they differ first in what a visitor sees by default. Google's reCAPTCHA v3 shows nothing: its documentation says "reCAPTCHA v3 will never interrupt your users" and returns a score where "1.0 is very likely a good interaction, 0.0 is very likely a bot", with 0.5 as the suggested threshold. reCAPTCHA v2 in its checkbox variant shows a checkbox that sometimes opens an image puzzle. hCaptcha shows a checkbox and image tasks, and Cloudflare Turnstile decides per request whether the visitor has to click.

ServiceOwnerWhat a visitor sees by defaultWhat the site receives
reCAPTCHA v3Googlenothinga score from 0.0 to 1.0
reCAPTCHA v2 checkboxGooglea checkbox, sometimes an image puzzlea token
hCaptchahCaptchaa checkbox and image tasks, set by the site's difficulty levela token
TurnstileCloudflaremanaged: a widget that passes by itself or asks for one click; invisible: nothinga token

What does each one ask of an agent?

Each one asked something different of our agent: Turnstile mostly nothing, reCAPTCHA v2 a picture puzzle, hCaptcha a timed set of image tasks. The agent worked every row through its own tools, and a row counts only if the page moved on.

Service and variantMeasured onWhat it askedTimeResult
Turnstile, managed5sim loginno interactiontoken in 1.7 secondspassed
Turnstile, invisibleCloudflare dashboard loginno interactiontoken in 2.9 seconds2 of 2 passed
Turnstile, forced checkboxCloudflare sign-upone click on the checkboxnot recordedfailed twice on 4 August, 2 of 2 passed on 14 August
reCAPTCHA v3Google's demono interactionnot recordedscore 0.9 in 2 runs, the same as ordinary Chrome
reCAPTCHA v2Google's demoan image puzzle, as for ordinary Chromefirst attempt3 correct tiles, passed
reCAPTCHA v2, audioGoogle's demoa spoken cliptranscribed in 1.3 secondspassed
hCaptchaa third-party deploymenttwo pages of image tasksabout 35 seconds per set4 of 4 passed
hCaptchahCaptcha's own demothe same image taskssameno token for correct answers
Source: Scalebrowser agent through its MCP tools, live widgets with production site keys, 4 to 14 August 2026
Turnstile in managed mode returned a token in 1.7 seconds without a click, reCAPTCHA v2 showed the same puzzle as to ordinary Chrome, and hCaptcha passed 4 of 4 on a third-party key while its own demo gave no token for the same answers.

Two rows prove less than they seem to. reCAPTCHA v3 returned a constant 0.9 on a fresh key without visitor history, to every browser we tried, so the score separates nothing on a demo. And reCAPTCHA v2 opened its puzzle for ordinary Chrome on the same line in the same minute, so the puzzle was the demo's answer to that line, not a verdict on our browser.

Which is harder for bots?

None of the three is harder for bots in general, because the site key and its settings decide more than the vendor does. The clearest case is hCaptcha: the same browser, in the same minute, on the same line, with correct answers, got a token on a third-party deployment and none on hCaptcha's own demo. hCaptcha's FAQ explains the lever: publishers choose between "Easy, Medium, Difficult, and Auto", and "Setting the value to Auto means the user will always be presented with a challenge". The owner of the key, not the brand on the widget, sets how hard the next step is.

Measured by effort, our agent's order was clear on these pages. Turnstile asked for nothing in 2 of its 3 variants, reCAPTCHA v2 asked for one puzzle a person also gets, and hCaptcha cost about 35 seconds per set of two pages inside a window that closes after about two minutes. All three challenge providers ended in a pass on at least one deployment, and each also has a deployment or a day where the same agent got nothing.

How does Cloudflare Turnstile differ from reCAPTCHA?

Cloudflare Turnstile differs from reCAPTCHA in that it replaces the puzzle with checks the browser runs by itself. Cloudflare's launch post of 28 September 2022 describes "a rotating suite of non-intrusive browser challenges based on telemetry and client behavior", among them "proof-of-work, proof-of-space, probing for web APIs". Its widget documentation adds that managed mode "automatically chooses between non-interactive or checkbox challenge based on visitor risk level". reCAPTCHA splits the same job in two products: v3 returns a score and leaves the decision to the site, v2 asks the visitor to prove it with a puzzle.

QuestionTurnstilereCAPTCHA v3reCAPTCHA v2 checkbox
Does the visitor act?only if a further check is neededneveralways the checkbox, sometimes a puzzle
Who decides the outcome?Cloudflare, per requestthe site, from the scoreGoogle, from the answer
Needs the Cloudflare network?nonono
Scalebrowser

Scalebrowser gives each agent its own isolated browser with a persistent identity, on your own machine, so a run stays signed in, handles the captcha and finishes without anyone watching it.

See how Scalebrowser does it

So which should you expect on a given site?

Expect the vendor's default only on a demo; on a real site, expect whatever the key's owner configured. The table sums up what our runs support.

SituationWhat to expectWhy
Turnstile on an ordinary logina token without a clickmanaged mode asked for no click on 5sim
Turnstile on the vendor's own sign-upa checkboxCloudflare asked for one on its own page
reCAPTCHA v2 on a demoan image puzzle, as for ordinary Chromeon the same line the demo gave it to both browsers
reCAPTCHA v3 anywherea score you never seeonly the site reads it
hCaptcha on a strict keyimage tasks, possibly no tokenthe publisher's difficulty level decides

Our verdict: a comparison that ranks the three by difficulty ranks their demos. For an agent, the useful questions are which variant a site runs, which the page shows before the first click, and how strict its key is, which only a run shows.

Run it on your own machine

Free to start, no card needed: one browser at a time with your own agents on your own sites.

Written by

Davide Valli

Founder, Scalebrowser

Builds Scalebrowser, the browser layer for AI agents that runs on your own machine. Measures every change a web page could observe against a real browser before it ships, and writes up the ones that turned out wrong.

All articles by Davide Valli