VLM-SubtleBench - Difference Captioning: leaderboard
Metric: Difference-captioning accuracy (%): share of generated captions of the difference between two similar images that a GPT-4o judge labels as matching the reference caption; higher is better. Source: arxiv.org. Saturation forecast: Around August 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 (Thinking) | 43 | #91 (GPT-5) |
| 2 | O3 | 38.1 | #121 |
| 3 | GPT-5 Chat | 37.2 | #233 |
| 4 | Gemini 2.5 Pro | 29.4 | #145 |
| 5 | Gemini 2.5 Flash | 25.9 | #237 |
| 6 | GPT-4o | 25.2 | #333 |
| 7 | Claude Sonnet 4 | 24.8 | #194 |
| 8 | Qwen 2.5 VL 72B Instruct | 24.3 | #364 |
| 9 | Qwen 2.5 VL 32B Instruct | 22.7 | #443 |
| 10 | Qwen 2.5 VL 7B Instruct | 12.7 | #643 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=vlm-subtlebench-difference-captioning · How It Works · Data refreshed daily, snapshot 2026-10-11.