UNICBench (Image): leaderboard
Metric: Exact-count hit rate (%): share of parsed numeric answers equal to the annotated count over UNICBench's 5,508 image counting questions (5,300 images; everyday objects, remote sensing, crowds, cells), temperature 0 (1.0 where a model fixes it), 4,096-token cap, GPT-5 and GPT-5-mini at minimal and o3 at low reasoning effort; unparsable responses are left out of the rate; higher is better. Source: arxiv.org. Saturation forecast: Around 2032. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 20.3 | #145 |
| 2 | O3 (Low) | 18.5 | #121 (O3) |
| 3 | Qwen 3 VL 30B A3B (Thinking) | 18.3 | #338 (Qwen 3 VL 30B A3B) |
| 4 | InternVL3-78B | 17.8 | #345 |
| 5 | Qwen 2.5 VL 72B Instruct | 17.4 | #364 |
| 6 | Qwen 3 VL 30B A3B Instruct | 17.3 | #365 |
| 7 | GPT-5 Mini (Minimal) | 17.2 | #176 (GPT-5 Mini) |
| 8 | O4 Mini | 17.1 | #172 |
| 9 | GPT-5 (Minimal) | 16.8 | #91 (GPT-5) |
| 10 | GPT-4o | 16.6 | #333 |
| 11 | Qwen 2.5 VL 7B Instruct | 16.6 | #643 |
| 12 | GLM-4.5V | 16 | #339 |
| 13 | GLM-4.1V-9B (Thinking) | 15.7 | #457 (GLM-4.1V-9B) |
| 14 | Gemini 2.5 Flash | 15.6 | #237 |
| 15 | Claude Sonnet 4 (20250514) | 14.9 | #211 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=unicbench-image · How It Works · Data refreshed daily, snapshot 2026-10-11.