HELM VHELM - Vibe Eval - Hard: leaderboard
VHELM (Stanford CRFM HELM for vision-language models, run v2.1.2) Vibe-Eval hard-difficulty prompts: open-ended everyday multimodal tasks with reference answers, each response rated 1-5 by a Prometheus-Vision judge; 100 prompts. Mean rating on the 1-5 scale, higher is better.
Metric: Prometheus Vision Rating (1-5). Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 48 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro (Preview 03-25) | 3.04 | #126 |
| 2 | O4 Mini (2025-04-16) | 2.65 | #183 |
| 3 | O1 (2024-12-17) | 2.52 | #157 |
| 4 | O3 (2025-04-16) | 2.49 | #125 |
| 5 | Claude 3.5 Sonnet (20241022) | 2.48 | #302 |
| 6 | GPT-4.1 (2025-04-14) | 2.46 | #257 |
| 7 | Claude 3.5 Sonnet (20240620) | 2.44 | #308 |
| 8 | Llama 4 Maverick Instruct FP8 | 2.42 | #424 |
| 9 | GPT-4.1 Mini | 2.39 | #363 |
| 10 | Gemini 2.0 Pro (Preview 02-05) | 2.38 | #318 |
| 11 | Gemini 1.5 Flash (002) | 2.37 | #587 |
| 12 | Claude 3 Opus (20240229) | 2.37 | #428 |
| 13 | Gemini 2.0 Flash (Preview) | 2.35 | #376 |
| 14 | GPT-4 Turbo | 2.29 | #417 |
| 15 | GPT-4.5 | 2.25 | #278 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-vhelm-vibe-eval-hard · How It Works · Data refreshed daily, snapshot 2026-11-05.