LongBench v2: leaderboard
LongBench v2: Measures long-context retrieval, needle finding, summarization, factual grounding, or retrieval-augmented generation quality.
Metric: Accuracy (%). Source: github.com. Source: longbench2.github.io. Saturation forecast: Around January 2027. 36 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 63.3 | #152 |
| 2 | Gemini 2.5 Flash | 62.1 | #245 |
| 3 | Qwen 3 235B A22B 2507 (Thinking) | 60.6 | #260 (Qwen 3 235B A22B 2507) |
| 4 | DeepSeek R1 | 58.3 | #251 |
| 5 | Qwen 3 235B A22B 2507 Instruct | 58.3 | #300 |
| 6 | O1 Preview | 57.7 | #193 |
| 7 | DeepSeek R1 0528 | 56.7 | #222 |
| 8 | MiniMax-Text-01 | 56.5 | #513 |
| 9 | Gemini 2.0 Flash (Thinking) | 56 | #342 (Gemini 2.0 Flash) |
| 10 | GPT-4o | 51.4 | #343 |
| 11 | Gemini 2.0 Flash | 51.1 | #342 |
| 12 | GLM-4.5 | 50.3 | #275 |
| 13 | Qwen 3 235B A22B | 50.1 | #313 |
| 14 | Qwen 3 30B A3B 2507 (Thinking) | 50.1 | #376 (Qwen 3 30B A3B 2507) |
| 15 | Qwen 3 32B | 49.2 | #443 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=longbench-v2 · How It Works · Data refreshed daily, snapshot 2026-10-20.