HallusionBench: leaderboard
HallusionBench evaluates multimodal large language models on visual illusion and hallucination-style image-text reasoning cases.
Metric: Overall Score. Source: github.com. Source: huggingface.co. Source: rank.opencompass.org.cn. Saturation forecast: Around April 2028. 284 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 65.2 | #98 |
| 2 | Gemini 2.5 Pro | 64.1 | #152 |
| 3 | GPT-5 Mini (2025-08-07) | 62.5 | #172 |
| 4 | GPT-5 Nano | 60.9 | #425 |
| 5 | GPT-4.5 | 60 | #274 |
| 6 | InternVL3-78B | 59.1 | #352 |
| 7 | Qwen 2 VL 72B | 58.7 | #460 |
| 8 | GPT-4.1 (2025-04-14) | 58.5 | #250 |
| 9 | Doubao-1.5-Pro | 58.5 | #287 |
| 10 | InternVL3-38B | 58.4 | #405 |
| 11 | Gemini 2.0 Flash | 58 | #342 |
| 12 | InternVL2.5-78B | 57.4 | #431 |
| 13 | GPT-4o ChatGPT | 57 | |
| 14 | Gemini 1.5 Pro (002) | 55.9 | #349 |
| 15 | InternVL3-14B | 55.9 | #505 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=hallusionbench · How It Works · Data refreshed daily, snapshot 2026-10-20.