CaptchaBench: leaderboard
CAPTCHA-solving benchmark from CaptchaBench (CaptchaMind paper, Wang et al., Alibaba Group with SUSTech and HIT Shenzhen, EMNLP 2026): a vision-language agent solves eight kinds of interactive CAPTCHA modelled on commercial services in a simulator, acting through click, drag and number-entry tool calls. The score is the mean success rate in percent over the eight tasks (200 generated test episodes each); higher is better. The authors' CaptchaMind models are fine-tuned on the benchmark's own training split.
Metric: Task success rate (%; mean of eight CAPTCHA tasks, 200 test episodes each). Source: arxiv.org. Saturation forecast: Around 2029. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4 | 47.4 | #397 |
| 2 | O3 | 46.9 | #224 |
| 3 | Claude 3.7 Sonnet | 43.9 | #455 |
| 4 | Qwen-VL-Max | 38.6 | #1226 |
| 5 | GPT-5 | 33.6 | #163 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=captchabench · How It Works · Data refreshed daily, snapshot 2026-11-12.