CaptchaBench - Image Select: leaderboard
CaptchaBench task (CaptchaMind paper, Wang et al., EMNLP 2026) in which the agent clicks every tile in an image grid that contains a named object, such as a bicycle, receiving feedback after each step. Success rate in percent over 200 generated test episodes, with one ReAct tool-call prompt for every model; higher is better.
Metric: Task success rate (%; 200 test episodes). Source: arxiv.org. Saturation forecast: Around 2029. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 52 | #135 |
| 2 | O3 | 46 | #224 |
| 3 | Claude Sonnet 4 | 43 | #397 |
| 4 | Qwen-VL-Max | 39.5 | #1226 |
| 5 | Claude 3.7 Sonnet | 38.5 | #455 |
| 6 | GPT-5 | 32.5 | #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-image-select · How It Works · Data refreshed daily, snapshot 2026-11-12.