FREAK - Counting: leaderboard
Metric: Accuracy (%) on the counting questions of FREAK's 1,799 questions (1,000 multiple-choice with cyclic option permutations and 799 free-form judged by GPT-5-mini) on photorealistic images edited to contradict commonsense; reasoning enabled where the model supports it; higher is better. Source: arxiv.org. Saturation forecast: Around 2033. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 23.98 | #145 |
| 2 | O4 Mini | 21.96 | #172 |
| 3 | O3 (High) | 21.14 | #121 (O3) |
| 4 | Gemini 2.5 Flash | 20.81 | #237 |
| 5 | InternVL3-78B | 20.12 | #345 |
| 6 | GPT-4.1 | 19.45 | #240 |
| 7 | GLM-4.5V | 19.41 | #339 |
| 8 | Phi-4 Multimodal Instruct | 18.89 | #896 |
| 9 | InternVL3-38B | 17.84 | #395 |
| 10 | Claude Sonnet 4 (Thinking) | 17.48 | #194 (Claude Sonnet 4) |
| 11 | Qwen 2.5 VL 72B Instruct | 16.84 | #364 |
| 12 | Qwen 2.5 VL 32B Instruct | 16.74 | #443 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=freak-counting · How It Works · Data refreshed daily, snapshot 2026-10-11.