VLM-SubtleBench - Temporal: leaderboard
Metric: Accuracy (%) on the temporal-difference two-image multiple-choice questions of VLM-SubtleBench's test split (pairs of highly similar images from natural, game, industrial, aerial, synthetic and medical sources); temperature 0.5; higher is better. Source: arxiv.org. Saturation forecast: Around September 2028. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 | 60.4 | #121 |
| 2 | GPT-5 (Thinking) | 60.2 | #91 (GPT-5) |
| 3 | Gemini 2.5 Pro | 57.6 | #145 |
| 4 | Gemini 2.5 Flash | 53.9 | #237 |
| 5 | GPT-5 Chat | 53.6 | #233 |
| 6 | GPT-4o | 52.7 | #333 |
| 7 | Qwen 2.5 VL 32B Instruct | 50.4 | #443 |
| 8 | Qwen 2.5 VL 7B Instruct | 50.2 | #643 |
| 9 | Qwen 2.5 VL 72B Instruct | 49.9 | #364 |
| 10 | Claude Sonnet 4 | 49.3 | #194 |
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-temporal · How It Works · Data refreshed daily, snapshot 2026-10-11.