These case studies document what our own production pipeline measured, how we measured it, and what integration lessons came out of the data. We publish measurements instead of testimonials: customer workloads are confidential, so the evidence here is our own benchmark data, collected on live traffic as described in the methodology.
Case Study 1 — GeeTest v4 Slider at Production Scale
400 production samples · June–August 2026 · GeeTest solver
What we measured
| Metric | Result |
|---|
| First-attempt success | 97% |
| Median solve time | 0.4s |
| p95 solve time | 0.9s |
What the data taught us
- Natural pixel size matters. Challenge images submitted at their natural resolution (not CSS-scaled screenshots) solved measurably better — scaling artifacts confuse piece localization.
- Human-like trajectories win. Slider trajectories in the 600–900ms range passed far more reliably than instant jumps or perfectly linear motion.
- Freshness is non-negotiable. Reusing a challenge image across attempts drops success sharply; every attempt must start from a freshly loaded challenge.
Integration details: GeeTest API docs.
Case Study 2 — Amazon AWS WAF Latency Distribution
500 production runs · January–August 2026 · Amazon WAF solver
What we measured
| Metric | Result |
|---|
| Success rate | 97% |
| Median solve time | 8s |
| p95 solve time | 14s |
| Maximum observed | 30s |
What the data taught us
- The tail is real. Median latency (8s) is less than half the p95 (14s). Clients that time out at 10s abandon roughly one solve in twenty.
- Set client timeouts at 30s. The slowest observed successful solve took 30s — anything shorter discards paid-for results that were about to arrive.
- Retries must be marked. Resubmitting with
is_retry=true prevents double billing when a previous attempt's status is uncertain.
Integration details: Amazon WAF solver guide.
Case Study 3 — Image/Text OCR Evaluation on a Labeled Set
1,000 labeled samples · August 2026 · Text captcha solver
What we measured
| Metric | Result |
|---|
| Accuracy (best conditions) | up to 99.9% |
| Median response time | 1.2s |
| p95 response time | 2.5s |
| Maximum observed | 10s |
What the data taught us
- Single-line crops perform best. Multi-line or heavily distorted images are the main accuracy drag; crop to the answer line before upload.
- Upscale small images ~2×. Very small captchas gain measurable accuracy from a simple 2× upscale before submission.
- Format discipline. JPG, PNG, and GIF up to 5MB are accepted by
/api/process.php; lossless formats (PNG/GIF) preserve edges better than re-compressed JPEGs.
Integration details: OCR endpoint reference.
Methodology & Reproducibility
Sample sizes, collection windows, and measurement definitions for every figure on this page are documented at /benchmarks. Benchmarks are re-measured quarterly and whenever a model version changes; material changes are announced on the changelog. Support status per captcha type is tracked on the capability matrix.