Which AI actually solves the captcha — cheapest?

Every candidate model run over the same saved puzzles, scored on the one thing that counts: clicking the right doodles in the right order. Ranked by solve-rate, then speed and cost. Captchas are rare — so what you actually pay to clear them is tiny. See the real monthly cost →

What is this?

We scrape the Smyths toy catalogue through Tor (free, anonymous IPs). Now and then the site throws up a security puzzle from hCaptcha — the doodle kind: a photo with a few faint hand-drawn icons scribbled on it, and a strip up top saying “select in this order.” You have to click the doodles in that exact sequence.

To get past it automatically you need a vision model that can (1) find the doodles and (2) put them in the right order. This page is the bake-off: same real puzzles, every model, head-to-head — so we can pick the one that solves it reliably for the least money. See where each model actually clicked →

Puzzles in set

—

Models tested

—

Best real solve-rate

—

Cheapest that clears target

—

solved ✓ = clicked every doodle in the right order — the only result that actually passes the captcha (the real score) · located = found all the doodles but maybe out of order (still fails) · icons hit = average share of doodles located · p50 = median time per puzzle · cost = USD per 1,000 captcha-solves (FREE→$$$), real bench tokens × real OpenRouter price. Hover it for the monthly cost.

💡 Cost is per 1,000 solves, not per scrape — the solver only runs when a captcha appears, which is rare. So the real monthly bill is tiny: see the Economics tab. A ~ means the unit price is estimated (model isn't a production pick).

What we've gathered — across every run

Products captured, pages fetched, Tor exits used, time spent, and the doodle captchas met and cleared — totalled across all runs.