measured, not claimed · 200-solve vendor benchmark
Stop paying more to fail more often.
Captchas are why your jobs come back half-empty. Legibl clears them behind the scenes, routes every solve to whoever is actually delivering that day, and shows you the receipts. Nobody on your side maintains anything.
What happens to a request that hits a captcha
waiting- 01Detect
- 02Route
- 03Solve
- 04Verify
- 05Resubmit
What stops being your problem
Three things. Prices in this market vary by 5.5× for the same commodity and the most expensive vendors are the least reliable, which is what makes the first two of them possible.
Jobs finish
The pipeline that quietly returned half a dataset returns all of it. When a vendor degrades we move your traffic before you notice, because we are measuring them continuously rather than trusting a published figure.
Nobody owns the captcha code
The workaround someone wrote last year, and everyone is now afraid to touch, comes out. It is not replaced with something heavier — nothing new joins your stack, and nothing new is yours to patch.
You see what it really cost
Not solves attempted, not a vendor's word for it — what the target accepted, per domain, in your own dashboard. Charged only when a solve comes back.
Priced per solve, published.
Prepaid credit, no subscription, no minimum and no volume negotiation. If nothing comes back, you are not charged for it.
recaptcha_v2
$0.80
per 1,000 solves
recaptcha_v2_invisible
$0.80
per 1,000 solves
turnstile
$1.20
per 1,000 solves
image_ocr
$0.60
per 1,000 solves
Image OCR is submitted, not detected
Image captchas are accepted by the API and priced above. Not yet handled by the SDKs' automatic path — you submit the image yourself.
What is not supported
hCaptcha, DataDome, Kasada, PerimeterX / HUMAN, FunCaptcha / Arkose are not supported. There is no detection and no task type for any of them. Four types, named exactly — because a type that is claimed but not implemented costs you a production incident to discover.
The claim we are not making yet
We measure cost per accepted solve, per target domain, and show it to you. We are not claiming it beats list-price shopping, because we have not demonstrated that: across 195 verified solves against a property we control, every token was accepted, so the two numbers were identical. That is a fact about an undefended target. Divergence is what a hostile target produces, and until we have measured one, the instrument is real and the conclusion is not.
What you may point it at
Systems you own, or systems whose operator has authorised you to access them. That is the line, it is the whole line, and it is set out with the specifics in the acceptable use policy. If you believe Legibl is being used against something you run, report it — no account required.
Point it at the job that keeps failing.
Prepay what you like, run it against a system you are authorised to access, and compare this week's numbers with last week's. That is the entire evaluation.