Benchmarks & Methodology

1. What This Page Is

This page is the single source of truth for every accuracy, speed, and uptime figure published on captchakings.com. Any percentage quoted on our homepage, solver pages, documentation, or blog refers to the measurement methods described here.

2. How We Measure Accuracy

• Ground-truth sampling. For each captcha type we draw random samples from production tasks and from labeled test sets whose correct answers are known in advance.
• Scoring rule. A task counts as correct only when the returned solution is accepted by the target system (token accepted, text matches ground truth, or the clicked elements match the labeled set).
• Per-type reporting. We publish accuracy per captcha type and variant. We do not publish a single sitewide accuracy number, because difficulty differs sharply between types.
• "Up to" figures. Where a page says "up to 99.9%" for text captchas, it refers to our standard distorted-text test set (section 3), not to every image a customer may submit. Heavily distorted, low-resolution, or unusual fonts score lower.

3. Measured Accuracy by Captcha Type

Captcha type	Variant	Sample size	Window	Measured accuracy
Tencent / Chinese	Slider, sequence click (顺序点击), spatial category	500+ production samples	Jun–Aug 2026	97%
GeeTest	v3 click, v4 slider	400 production samples	Jun–Aug 2026	97%
xCaptcha	Binary button classification	500 production samples	Jun–Aug 2026	96%
MTCaptcha	Image-to-text token	300 production samples	Jun–Aug 2026	98%
Amazon AWS WAF	captcha_voucher token	500 production samples	Jan–Aug 2026	97%
Image / Text OCR	Standard distorted text test set	1,000 labeled images	Aug 2026	up to 99.9%

Production accuracy for customer-submitted images varies with distortion, resolution, and font. The figure above reflects our standard test set; your results may differ. Failed tasks are not billed beyond the attempt price shown on our pricing page.

4. How We Measure Response Time

Response time is measured server-side from the moment a request is accepted to the moment the solution is ready for retrieval. It excludes your network latency to our API. We report the median, the 95th percentile (p95), and the maximum observed in the window:

Captcha type	Median	p95	Max observed
Tencent / Chinese	0.17s	0.35s	2s
GeeTest v3/v4	0.4s	0.9s	3s
xCaptcha	0.8s	1.5s	5s
MTCaptcha	2.1s	4s	10s
Amazon AWS WAF	8s	14s	30s
Image / Text OCR	1.2s	2.5s	10s

"Processing latency" quoted on some solver pages (for example "<200ms" for Tencent) refers to model inference time only, which is one component of the end-to-end figures above.

5. How We Measure Uptime

• Method. Synthetic probes against our public API endpoints every 60 seconds from two monitoring regions. A probe fails when the endpoint does not respond or returns a server error.
• Window. Calendar month, aligned with our Service Level Agreement.
• Exclusions. Scheduled maintenance announced at least 24 hours in advance, as defined in the SLA.
• Target. 99.9% monthly uptime. Remedies when we miss the target are defined in the SLA.

6. Internal vs. Independent Measurement

All numbers on this page are produced by our own instrumentation. They are not customer-reported averages and not third-party benchmarks. We publish the method, sample sizes, and windows so you can reproduce comparable tests against the live API using your free balance.

7. Update Policy

Figures are re-measured quarterly and whenever a model version changes. Material changes are announced on the changelog and AI updates pages. The "last verified" date at the top of this page reflects the most recent review of every number shown here. See also our capability matrix for per-type support status.