GapEval - Reasoning (Understanding): leaderboard
GapEval (Wang et al., 2026) poses bidirectional multimodal questions that a model can answer in text or with a generated image; these boards score the text answers. Reasoning: image selection from diagrams and game layouts, knowledge questions rendered as images, and real-world and logical reasoning problems. Accuracy in percent, judged correct or incorrect by GPT-5-mini against a reference answer and averaged over ten sampled runs; higher is better.
Metric: Accuracy (%; image selection, knowledge selection, real-world and logical reasoning problems; text answers to GapEval's bidirectional questions, judged correct or incorrect by GPT-5-mini against reference answers, mean of ten sampled runs). Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 77.05 | #452 |
| 2 | GPT-5 | 74.74 | #163 |
| 3 | GPT-5 Mini | 73.8 | #349 |
| 4 | Qwen 3 VL 235B A22B | 60.2 | #477 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=gapeval-reasoning-understanding · How It Works · Data refreshed daily, snapshot 2026-11-12.