DiffSpot: leaderboard
Fine-grained visual difference benchmark on web-page screenshots (DiffSpot, Zhang et al., WeChat AI, Tencent, 2026): each pair differs by one CSS property change on one element, drawn from 13 operators in three magnitude tiers, or not at all, and the model must describe what changed. A gpt-oss-120b judge credits an answer that names the recorded change, or that reports no change on the 500 unchanged controls; the score is accuracy in percent over all 4,400 pairs, zero-shot with one prompt for all models; higher is better.
Metric: Accuracy (%; (TP+TN)/4,400 over 3,900 changed pairs and 500 unchanged controls, gpt-oss-120b judge). Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 47.2 | #91 |
| 2 | Kimi K2.5 | 42.2 | #252 |
| 3 | Gemini 3 Flash | 40.9 | #159 |
| 4 | Claude Opus 4.7 | 38.9 | #75 |
| 5 | GPT-5.4 | 38.3 | #121 |
| 6 | Qwen 3.5 397B A17B | 37.6 | #242 |
| 7 | Qwen 3 VL 235B A22B (Thinking) | 28.3 | #451 |
| 8 | GLM-4.6V | 21.2 | #698 |
| 9 | Qwen 3 VL 30B A3B Instruct | 17.6 | #870 |
| 10 | Qwen 3 VL 30B A3B (Thinking) | 17.5 | #710 |
| 11 | Qwen 3 VL 235B A22B Instruct | 15.9 | #521 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=diffspot · How It Works · Data refreshed daily, snapshot 2026-11-12.