Big-Bench Hard — leaderboard
23 challenging tasks from BIG-Bench where prior language models failed to outperform average human raters. Tests algorithmic reasoning, language understanding, and world knowledge.
Metric: Average (%). Source: epoch.ai. Status: saturation imminent. 50 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 1.5 Pro (001) | 89.2 |
| 2 | DeepSeek V3 | 87.5 |
| 3 | Llama 3.1 405B | 82.9 |
| 4 | Phi-3-medium-128k-instruct | 81.4 |
| 5 | Qwen 2.5 72B | 79.8 |
| 6 | Phi-3-small-8k-instruct | 79.1 |
| 7 | DeepSeek V2 | 78.8 |
| 8 | GPT-4 (0613) | 75.12 |
| 9 | Phi-3 Mini 4K Instruct | 71.7 |
| 10 | Yi 34B (Chat) | 71.7 |
| 11 | StableBeluga2 | 69.3 |
| 12 | Llama 2 70B Base | 64.9 |
| 13 | GPT-3.5 Turbo (0613) | 61.59 |
| 14 | Phi-2 | 59.4 |
| 15 | Nemotron-4 15B | 58.7 |
Interactive version: theaggregate.ai/benchmark?slug=big-bench-hard · How the rankings work · Data refreshed daily, snapshot 2026-07-22.