MRCR v2 (8-needle, Context Arena) - 128K pointwise
General QA
Context Arena's mean MRCR v2 8-needle string similarity over the 103 tests in its 128K bin (bin key 131,072 tokens). It is a pointwise reading from the same evaluator runs as the family's 128K AUC primary, and every Context Arena entry reports it.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 92.4 |
| Qwen3.8-Max | 92.3 |
| Claude Opus 5 | 91.3 |
| Gemini 3.7 Flash | 88.6 |
| GPT-5.5 | 85.8 |
| Qwen3.8 27B | 81.8 |
| GPT-5.6 Terra | 80.2 |
| Gemini 3.6 Flash | 77.3 |
| Opus 4.8 | 75.2 |
| Sonnet 4.6 | 72.6 |
| Opus 4.6 | 72.5 |
| Qwen3.6-Max-Preview | 71.6 |
| GLM-5.3 | 69.3 |
| GPT-5.4 | 67.8 |
| GLM-5.3 Flash | 62.9 |
| DeepSeek-V4-Flash-0731 | 62.9 |
| DeepSeek V4 Pro 0813 | 60.4 |
| Muse Spark 1.2 | 60.3 |
| Qwen3.7 Flash | 59.3 |
| Qwen3.6 35B A3B | 59.3 |
| GPT-5.6 Luna | 58.4 |
| Gemini 3.1 Pro Preview | 57.6 |
| Gemini 3.5 Flash | 57.5 |
| Qwen3.6 27B | 56.6 |
| Qwen3.6 Plus (2026-04-02) | 56.6 |
| Gemini 3.5 Flash-Lite | 55.5 |
| Sonnet 5 | 52.9 |
| Grok 4.6 | 52.6 |
| Ring 2.6 1T | 52.1 |
| Muse Spark 1.1 | 51.8 |
| GLM-5.2 | 51.6 |
| DeepSeek-V4-Pro | 48.4 |
| Ling 3.0 Flash | 45.4 |
| Qwen3.5 Plus (2026-02-15) | 44.3 |
| DeepSeek-V4-Flash | 43.9 |
| Qwen3.5-27B | 42.2 |
| Qwen3.5 397B A17B | 42.0 |
| Qwen3.7-Max (source unspecified) | 41.4 |
| Kimi K3 | 39.5 |
| Kimi K2.6 | 39.5 |