GEB-Bench - Average (Vision Models): leaderboard
GEB-Bench (arXiv 2608.04111v2, Tables 2-3): mean accuracy over all eight tasks for vision-capable models, recognizing 25 abstract structural motifs (fixed points, dualities, cycles and others) from scenes, skeleton diagrams, stories and theorems and matching them across voices. Percent; higher is better. Text-only models are excluded.
Metric: Macro-average accuracy over all eight tasks (%). Source: arxiv.org. Saturation forecast: Around February 2027. 27 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 82.4 | #33 |
| 2 | Gemini 3.1 Pro (Preview) | 82.1 | #64 |
| 3 | GPT-5.2 | 76.9 | #115 |
| 4 | Kimi K2.5 | 76.9 | #152 |
| 5 | Gemma 4 31B | 72.1 | #212 |
| 6 | Claude Opus 4.8 | 69.8 | #39 |
| 7 | GLM-4.6V | 63 | #330 |
| 8 | Llama 4 Scout | 53.7 | #648 |
| 9 | InternVL3-8B | 45.5 | #624 |
| 10 | Qwen 2 VL 7B | 36.2 | #743 |
| 11 | ERNIE 4.5 VL 28B A3B | 35.6 | |
| 12 | Pixtral-12B | 34.2 | #805 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=geb-bench-average-vision-models · How It Works · Data refreshed daily, snapshot 2026-11-05.