R3-Bench: leaderboard
R3-Bench (2026) evaluates reflective visual generation: given a prompt and an image, a model judges whether they match and explains the discrepancy. Reflective Verdict Score in percent over 670 expert-annotated pairs in eight error categories, a misaligned pair counting only when the explanation matches the expert diagnosis according to an LLM judge; higher is better.
Metric: Reflective Verdict Score times 100 (%) averaged over eight error categories: an aligned image-prompt pair counts when the model says it matches; a misaligned one counts only when the model rejects it and an LLM judge finds its explanation equivalent to the expert diagnosis; 670 expert-annotated pairs; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 87 | #145 |
| 2 | Nano Banana | 84 | |
| 3 | GPT-5.2 | 80 | #185 |
| 4 | Qwen 3 VL 8B | 80 | #917 |
| 5 | GPT Image 1 | 79 | |
| 6 | GPT-4o | 76 | #751 |
| 7 | Qwen 2.5 VL 7B Instruct | 72 | #1902 |
| 8 | BAGEL-7B-MoT | 56 | #4619 |
| 9 | OmniGen2 | 51 | #4579 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=r3-bench · How It Works · Data refreshed daily, snapshot 2026-11-21.