LivingScreen: leaderboard
LivingScreen (Beijing Institute of Technology, 2026) runs multimodal GUI agents on a browser replica of a Chinese short-video app whose videos keep playing, so the agent must choose what to watch and for how long while it swipes, seeks and operates the feed; 499 annotated tasks, 30 steps per episode. The board is the success rate in percent, averaged over the three task tiers; higher is better.
Metric: Success rate (%), mean of the three task tiers; LivingScreen: 499 Chinese short-video tasks in a browser replica of a short-video app (1,528 real videos, about 5.6 per feed); a single-agent loop issues GUI actions plus watch and wait primitives from screenshots or recorded clips, 30-step cap, temperature 0.6, high reasoning effort; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.5 Flash (High) | 69.3 | #101 (Gemini 3.5 Flash) |
| 2 | Gemini 3.1 Pro (Preview) (High) | 66.2 | #94 (Gemini 3.1 Pro (Preview)) |
| 3 | Seed 2.0 Pro (High) | 64.8 | #211 (Seed 2.0 Pro) |
| 4 | Claude Opus 4.6 (High) | 45.6 | #102 (Claude Opus 4.6) |
| 5 | Qwen 3.6 Plus (High) | 44.9 | #222 (Qwen 3.6 Plus) |
| 6 | GLM-5V Turbo | 42.7 | #344 |
| 7 | Claude Sonnet 4.6 (High) | 38.7 | #161 (Claude Sonnet 4.6) |
| 8 | Kimi K2.5 (Thinking) | 36.8 | #256 (Kimi K2.5) |
| 9 | GPT-5.5 (High) | 29.8 | #49 (GPT-5.5) |
| 10 | GPT-5.4 (High) | 25.2 | #123 (GPT-5.4) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=livingscreen · How It Works · Data refreshed daily, snapshot 2026-11-13.