CAPTCHA scoreboard: agents, solvers and widgets, measured

The scoreboard answers one question for every audience: what does it take, today, to get past each kind of human check, and how fast is that changing? Site owners learn whether their CAPTCHA still holds. Agent builders learn what will stop them. Researchers get a re-test of published designs.

Latest results from the measurement page

First results after the 7-day warm-up, ~2026-11-10. Our domain is new, and Google warns that reCAPTCHA scores in a site's first week may differ from later ones, so nothing is published before the warm-up ends. When the measurement service publishes its summary, this box fills in by itself from https://measure.captcharena.org/api/summary.

Data source: https://measure.captcharena.org. Raw verdicts come from each vendor's own server check.

What will be on it

Lane	What is measured	Status (2026-11-03)
Real widgets	Server-accepted rate per agent setup on reCAPTCHA v2 and v3, Turnstile and self-hosted open CAPTCHAs on our own domain; the v3 score distribution per setup and network type	Measurement page live; warm-up until ~2026-11-10
Practice puzzles	Accuracy, time and cost per solve on open research puzzles, fresh instances each cycle	Coming
Behaviour	What an agent does when a CAPTCHA appears in the middle of a task: solve it, hand off to its human, refuse, call a paid solver, or misrepresent itself	Coming
Solving services	Acceptance rate, time and effective price of commercial services, bought as an ordinary customer at list price	Coming
Hosted browsers	How often hosted agent browsers are challenged and accepted	Coming
Half-life	For each published "AI-proof" design: the months from launch until agents pass it	Coming; method

Why each lane exists and how it is run: Scoreboard method. Lane-by-lane detail: Scoreboard lanes.

Why another scoreboard

The existing ones each cover one slice, and most are run by a party with something to sell (checked 2026-11-01):

Existing source	What it covers	What it lacks
CaptchaTheCat	24 solving services, daily	No named operator, no published method, paid-visibility pitch
HasData benchmark	5 solving services, 6 types, 50 challenges each	One-off article by a seller of scraping APIs; test pages not disclosed
Scrapeway	7 scraping APIs against 11 protected sites	Operator not named; cadence stated inconsistently
Open CaptchaWorld	AI agents on logic puzzles	Frozen at publication (May 2025)

None puts agents, solving services and real widgets side by side, and none publishes real server verdicts continuously. The closest precedent is a research paper, Broken Gates, which put real widgets on its own domains and checked every token with the vendor's server, once.