LongBench Pro - Hard - Thinking: leaderboard
Evaluates long-context understanding across English and Chinese tasks at four difficulty levels. Non-thinking and thinking results use separate published columns; missing settings remain absent. This series reports Hard - Thinking using Score (%). Source: the original paper result table.
Metric: Score (%). Source: arxiv.org. Saturation forecast: Around February 2027. 46 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 81.03 | #159 |
| 2 | GPT-5 | 78.74 | #105 |
| 3 | Claude Sonnet 4 | 74.72 | #207 |
| 4 | Gemini 2.5 Flash | 72.19 | #255 |
| 5 | DeepSeek V3.2 | 67.73 | #210 |
| 6 | Qwen 3 235B A22B 2507 (Thinking) | 67.1 | #263 |
| 7 | DeepSeek V3.1 | 62.22 | #275 |
| 8 | Qwen 3 Next 80B A3B (Thinking) | 61.46 | #312 |
| 9 | Qwen 3 235B A22B 2507 Instruct | 58.6 | #315 |
| 10 | Claude 3.7 Sonnet | 56.58 | #259 |
| 11 | Qwen 3 Next 80B A3B Instruct | 54.74 | #381 |
| 12 | DeepSeek R1 0528 | 53.68 | #232 |
| 13 | DeepSeek R1 | 53.39 | #262 |
| 14 | Qwen 3 30B A3B 2507 (Thinking) | 52.76 | #379 |
| 15 | GLM-4.6 | 48.92 | #265 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=longbench-pro-hard-thinking · How It Works · Data refreshed daily, snapshot 2026-11-05.