HELM VHELM - Bingo - T2I: leaderboard
VHELM (Stanford CRFM HELM for vision-language models, run v2.1.2) Bingo T2I split: open-ended questions designed to expose hallucination from text-to-image interference, where the prompt makes misleading claims about the image, each response rated 1-5 by a Prometheus-Vision judge; 128 questions. 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 | O3 (2025-04-16) | 3.72 | #125 |
| 2 | Gemini 2.5 Pro (Preview 03-25) | 3.64 | #126 |
| 3 | O4 Mini (2025-04-16) | 3.62 | #183 |
| 4 | GPT-4o Mini (2024-07-18) | 3.61 | #568 |
| 5 | Claude 3 Opus (20240229) | 3.58 | #428 |
| 6 | GPT-4o (2024-08-06) | 3.53 | #343 |
| 7 | GPT-4.1 Mini | 3.48 | #363 |
| 8 | Llama 4 Scout Instruct | 3.44 | #547 |
| 9 | Gemini 2.0 Pro (Preview 02-05) | 3.44 | #318 |
| 10 | GPT-4 Turbo | 3.42 | #417 |
| 11 | GPT-4.1 (2025-04-14) | 3.38 | #257 |
| 12 | Claude 3.5 Sonnet (20240620) | 3.34 | #308 |
| 13 | O1 (2024-12-17) | 3.34 | #157 |
| 14 | Gemini 2.0 Flash Lite | 3.34 | #455 |
| 15 | Llama 4 Maverick Instruct FP8 | 3.31 | #424 |
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-bingo-t2i · How It Works · Data refreshed daily, snapshot 2026-11-05.