LongBench Pro - Extreme - Non-Thinking: leaderboard
Metric: Score (%). Source: arxiv.org. 34 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Flash | 44.26 |
| 2 | Claude Sonnet 4 | 42.92 |
| 3 | Qwen 3 235B A22B 2507 Instruct | 42.07 |
| 4 | DeepSeek V3.1 | 41.07 |
| 5 | DeepSeek V3.2 | 40.45 |
| 6 | DeepSeek V3 (0324) | 40.4 |
| 7 | Kimi K2 0905 | 39.61 |
| 8 | Qwen 3 Next 80B A3B Instruct | 39.39 |
| 9 | Claude 3.7 Sonnet | 37.31 |
| 10 | GLM-4.6 | 37.3 |
| 11 | GPT-4o | 36.3 |
| 12 | Qwen 3 30B A3B 2507 Instruct | 35.44 |
| 13 | GLM-4.5 | 35.06 |
| 14 | Ministral-3-14B-Instruct-2512 | 33.85 |
| 15 | MiniMax-Text-01 | 33.57 |
Interactive version: theaggregate.ai/benchmark?slug=longbench-pro-extreme-non-thinking · How It Works · Data refreshed daily, snapshot 2026-09-19.