HELM VHELM - Toxicity: leaderboard
The "Toxicity" subtask of HELM VHELM. Stanford HELM's VHELM project, the holistic evaluation of vision-language models: image-grounded question answering, OCR, chart and diagram reading, spatial reasoning, robustness, bias, fairness and toxicity scenarios scored under one ha...
Metric: Mean win rate. Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 48 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O4 Mini (2025-04-16) | 95.74 | #181 |
| 2 | GPT-4.5 | 91.49 | #276 |
| 3 | O3 (2025-04-16) | 89.36 | #123 |
| 4 | O1 (2024-12-17) | 87.23 | #152 |
| 5 | Gemini 2.5 Pro (Preview 03-25) | 85.11 | #124 |
| 6 | Llama 4 Scout Instruct | 82.98 | #542 |
| 7 | Llama 4 Maverick Instruct FP8 | 76.6 | #412 |
| 8 | Gemini 2.0 Flash | 74.47 | #344 |
| 9 | GPT-4.1 Mini | 70.21 | #351 |
| 10 | Claude 3.5 Sonnet (20241022) | 67.02 | #291 |
| 11 | Gemini 2.0 Pro (Preview 02-05) | 67.02 | #305 |
| 12 | Gemini 2.0 Flash Lite | 63.83 | #453 |
| 13 | Gemini 1.5 Pro (002) | 59.57 | #350 |
| 14 | GPT-4o (2024-11-20) | 55.32 | #378 |
| 15 | GPT-4 Turbo | 51.06 | #406 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-vhelm-toxicity · How It Works · Data refreshed daily, snapshot 2026-10-23.