ADeptS-Bench (Desktop): leaderboard
ADeptS-Bench (2026) tests computer-use agents on single decision points in mobile and desktop screenshots. In the safety stream each task has a benign and a malicious variant whose interface hides a threat such as a phishing overlay or injected prompt; the score over 501 desktop task pairs is the harmonic mean of benign task success and one minus attack success, in percent; higher is better.
Metric: ADeptS-Bench score (%; harmonic mean of benign task success and one minus attack success, so capability and safety must both be high; 501 desktop task pairs across about 150 applications; paired benign and malicious GUI tasks with visual threats (phishing overlays, misleading buttons, prompt injection), graded on the next action against ground-truth boxes; mean of 10 runs at temperature 0.1). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 76 | #94 |
| 2 | Claude Opus 4.7 | 75.4 | #79 |
| 3 | GPT-5.4 | 66.1 | #123 |
| 4 | Qwen 3 VL 235B A22B | 37.2 | #479 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=adepts-bench-desktop · How It Works · Data refreshed daily, snapshot 2026-11-13.