OpenCompass Subjective - Math (CompassBench 2404): leaderboard
OpenCompass CompassBench 2404 edition (file dated 2024-05-09): subjective open-ended evaluation, math score. CompassBench is OpenCompass's periodic evaluation of chat models on a private question set; each edition is a separate board. Score 0-100, higher is better.
Metric: Score (%). Source: rank.opencompass.org.cn. Saturation forecast: Estimated already saturated. 51 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 Turbo | 56.2 | #417 |
| 2 | GPT-4 Turbo (Preview) | 50 | #416 |
| 3 | GPT-4o (2024-05-13) | 49.2 | #337 |
| 4 | Claude 3 Opus | 47.8 | #565 |
| 5 | GLM-4 | 42.2 | #647 |
| 6 | Qwen 1.5 110B Chat | 41.5 | #693 |
| 7 | Llama 3 70B Instruct | 38.5 | #640 |
| 8 | Moonshot v1 8K | 36 | #622 |
| 9 | Mistral Large | 35.5 | #713 |
| 10 | Qwen 1.5 72B Chat | 34.3 | #782 |
| 11 | DeepSeek V2 Chat | 33.8 | #572 |
| 12 | Qwen-72B-Chat | 32.8 | #835 |
| 13 | Qwen 1.5 14B Chat | 27.5 | #976 |
| 14 | Qwen-14B-Chat | 25.5 | #1161 |
| 15 | Mixtral 8x7B Instruct | 22 | #1085 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=opencompass-subjective-math-compassbench-2404 · How It Works · Data refreshed daily, snapshot 2026-11-05.