HELM Lite - MATH Level 1: leaderboard
HELM Lite v1.13.0 MATH with chain-of-thought prompting, restricted to Level 1, the easiest of the five MATH difficulty levels: the mean over seven subject runs (437 problems), up to 8-shot, scored by equivalence of the final answer to the reference. Percent, 0-100, higher is better; not comparable with full-MATH scores.
Metric: Equivalent (CoT) (%). Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 91 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 1.5 Pro (002) | 91.99 | #368 |
| 2 | DeepSeek V3 | 91.21 | #329 |
| 3 | Gemini 1.5 Flash (002) | 90.77 | #587 |
| 4 | Claude 3.5 Sonnet (20241022) | 90.37 | #302 |
| 5 | Gemini 2.0 Flash (Preview) | 90.13 | #376 |
| 6 | Qwen 2.5 72B Instruct | 88.39 | #469 |
| 7 | Claude 3.5 Haiku (20241022) | 87.17 | #590 |
| 8 | GPT-4 Turbo (Preview) | 85.69 | #416 |
| 9 | GPT-4o (2024-08-06) | 85.25 | #343 |
| 10 | GPT-4 Turbo | 83.29 | #417 |
| 11 | GPT-4o (2024-05-13) | 82.89 | #337 |
| 12 | Llama 3.1 405B Instruct | 82.7 | #470 |
| 13 | Gemini 1.5 Pro (001) | 82.47 | #465 |
| 14 | Nova Pro | 82.14 | #594 |
| 15 | Claude 3.5 Sonnet (20240620) | 81.28 | #308 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-lite-math-level-1 · How It Works · Data refreshed daily, snapshot 2026-11-05.