Benchmarks · long-context
Context Arena — MRCR v2 (8-needle)
The model must find and reproduce the right one of eight near-identical earlier replies hidden in a long conversation, averaged over context lengths from 8k to 128k tokens. Maintained by Context Arena. Score: Cumulative average score up to 128k tokens (unweighted mean of the 8k–128k bins), percent.
64 models · 66 results · data as of 2026-10-29 · contextarena.ai ↗
Leaderboard
GPT-5.6 Sol max
All results
66 rows| # | Model | Effort | Score |
|---|---|---|---|
| 1 | GPT-5.6 Sol | max | 97.9% |
| 2 | Opus 5 | max | 97.8% |
| 3 | Qwen 3.8 Max | xhigh | 96.9% |
| 4 | Gemini 3.7 Flash | max | 96.0% |
| 5 | Qwen 3.8 Max | medium | 95.2% |
| 6 | Qwen 3.8 27B | max | 94.4% |
| 7 | Qwen 3.8 Max | low | 93.1% |
| 8 | GPT-5.6 Terra | max | 92.6% |
| 9 | GPT-5.5 | xhigh | 91.2% |
| 10 | Opus 4.8 | max | 90.5% |
| 11 | Gemini 3.6 Flash | xhigh | 89.3% |
| 12 | GLM 5.3 | max | 88.8% |
| 13 | Qwen 3.6 Max | max | 88.0% |
| 14 | DeepSeek V4 Pro 0813 | max | 86.9% |
| 15 | GPT-5.4 | xhigh | 86.8% |
| 16 | Gemini 3.1 Pro | high | 85.1% |
| 17 | Qwen 3.6 35B-A3B | max | 83.6% |
| 18 | Sonnet 4.6 | max | 83.1% |
| 19 | Qwen 3.6 27B | max | 82.3% |
| 20 | Opus 4.6 | max | 82.2% |
| 21 | GPT-5.6 Luna | max | 82.1% |
| 22 | Grok 4.6 | max | 81.5% |
| 23 | Qwen 3.7 Flash | max | 81.2% |
| 24 | Muse Spark 1.2 | max | 80.6% |
| 25 | Sonnet 5 | max | 80.0% |
| 26 | GLM 5.3 Flash | max | 79.5% |
| 27 | Gemini 3.5 Flash | xhigh | 78.2% |
| 28 | DeepSeek V4 Flash 0731 | max | 77.9% |
| 29 | DeepSeek V4 Pro | max | 77.4% |
| 30 | Qwen 3.5 397B-A17B | max | 76.6% |