DISBench: leaderboard
Visual question answering with the ImageSeeker image-search agent and an explicitly identified vision-language backbone. Overall, exact-match, within-image and across-image results measure different parts of answer correctness. This board reports DISBench using Answer F1 (%).
Metric: Answer F1 (%). Source: huggingface.co. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 (20251101) | 55 | #81 |
| 2 | Gemini 3 Pro (Preview) | 47.9 | #71 |
| 3 | Claude Sonnet 4.5 | 43.8 | #148 |
| 4 | Gemini 3 Flash (Preview) | 36.8 | #90 |
| 5 | GPT-5.2 | 30 | #115 |
| 6 | Qwen 3 VL 235B A22B Instruct | 27.2 | #280 |
| 7 | GLM-4.6V | 26.1 | #330 |
| 8 | GPT-4o | 21.8 | #357 |
| 9 | Qwen 3 VL 32B Instruct | 20.9 | #290 |
| 10 | Qwen 3 VL 235B A22B (Thinking) | 20.4 | #236 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=disbench · How It Works · Data refreshed daily, snapshot 2026-11-05.