HELM (Stanford): leaderboard
Stanford HELM evaluation of language models across 42 scenarios and 7 metrics: accuracy, fairness, bias, toxicity, efficiency, robustness, and calibration.
Metric: Mean Win Rate (%). Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 91 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-05-13) | 93.76 | #325 |
| 2 | GPT-4o (2024-08-06) | 92.76 | #336 |
| 3 | DeepSeek V3 | 90.83 | #321 |
| 4 | Claude 3.5 Sonnet (20240620) | 88.5 | #299 |
| 5 | Nova Pro | 88.48 | #569 |
| 6 | GPT-4 (0613) | 86.71 | #491 |
| 7 | GPT-4 Turbo | 86.41 | #404 |
| 8 | Llama 3.1 405B Instruct | 85.39 | #457 |
| 9 | Claude 3.5 Sonnet (20241022) | 84.61 | #291 |
| 10 | Gemini 1.5 Pro (002) | 84.19 | #349 |
| 11 | Llama 3.2 90B Vision Instruct | 81.94 | #622 |
| 12 | Gemini 2.0 Flash (Preview) | 81.27 | #370 |
| 13 | Llama 3.3 70B Instruct | 81.22 | #534 |
| 14 | Llama 3.1 70B Instruct | 80.83 | #561 |
| 15 | Llama 3 70B | 79.29 | #631 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-stanford · How It Works · Data refreshed daily, snapshot 2026-10-20.