How the Hindi accuracy number was measured
2026-02-13 · Hindi captchas
175 live captchas, held out from training, all answered correctly.
The 100% figure is not a training-set score. The 175 images were live portal captchas the model had not been trained on. Every one was answered correctly.
That does not mean the next portal font is guaranteed. It means the measured set, which is the kind of sum these portals draw, was solved. A new colour or a new spelling can still miss, and the miss should be unreliable rather than a wrong charge.
Reliable is a lower bar here (0.5) than for plain text (0.8), because the snap onto a known word removes a class of error the text engine still has. The flag still means the service is willing to charge.
Keep a few of your own portal samples aside and run them before you depend on a new flow. The public samples on the site are for the shape of the request, not a promise about your portal.
Ready to call it? Start with the quickstart, the API guide, or pricing.