Numeric accuracy and the reliable subset
2026-02-28 · Numeric captchas
The headline accuracy and the accuracy among reliable answers are different numbers. Use the second one to decide.
88.3% of 60 live numeric captchas matched the human text. Among the ones the service marked reliable, 98.1% matched. The gap is the point of the flag.
If you submit every text and ignore the flag, you inherit the lower number and you also pay only for the reliable slice. The unreliable texts were free and more often wrong. Submitting them does not save money. It fails the form.
Your own portal can sit above or below that 60-image set. Measure it with keys and the dashboard usage chart before you promise a rate internally.
The usage chart groups by day and engine. Numeric captchas count under ocr.
Ready to call it? Start with the quickstart, the API guide, or pricing.