# Context Arena (MRCR)

Hides several near-identical items in a very long conversation and asks the AI to retrieve a specific one, checking whether it can find the right detail in a huge pile of text.

Category: long-context. Type: leaderboard. Status: active.
Metric: Average retrieval score (%) on the 8-needle task for conversations of 64k-128k tokens (higher is better).
Source: Context Arena — https://contextarena.ai/ (data from https://contextarena.ai/api/leaderboard?needles=8), retrieved 2026-11-17, license: not stated.

| Rank | Model [reasoning mode] | Score at 64k-128k tokens (%) | Overall score (%) |
|---|---|---|---|
| 1 | openai/gpt-5.6-sol [max] | 92.4 | 85.5 |
| 2 | qwen/qwen3.8-max [xhigh] | 92.3 | 73.7 |
| 3 | anthropic/claude-opus-5 [max] | 91.3 | 77.3 |
| 4 | google/gemini-3.7-flash [high] | 88.6 | 83.0 |
| 5 | qwen/qwen3.8-max [medium] | 87.3 | 72.1 |
| 6 | openai/gpt-5.5 [xhigh] | 85.8 | 79.8 |
| 7 | openai/gpt-5.5 [medium] | 83.0 | 78.6 |
| 8 | google/gemini-3.7-flash [medium] | 82.1 | 79.6 |
| 9 | qwen/qwen3.8-27b [low] | 81.8 | 92.8 |
| 10 | qwen/qwen3.8-max [low] | 81.5 | 71.6 |
| 11 | openai/gpt-5.6-terra [max] | 80.2 | 76.6 |
| 12 | openai/gpt-5.5 [high] | 79.3 | 79.0 |
| 13 | qwen/qwen3.8-27b [medium] | 78.9 | 94.0 |
| 14 | google/gemini-3.7-flash [low] | 78.3 | 74.5 |
| 15 | google/gemini-3.6-flash [high] | 77.3 | 73.3 |
| 16 | anthropic/claude-opus-4.8 [max] | 75.2 | 72.6 |
| 17 | google/gemini-3.6-flash [medium] | 74.5 | 69.9 |
| 18 | qwen/qwen3.8-27b [xhigh] | 74.3 | 90.2 |
| 19 | anthropic/claude-sonnet-4.6 [low] | 72.6 | 70.4 |
| 20 | anthropic/claude-opus-4.6 [medium] | 72.5 | 73.1 |
| 21 | anthropic/claude-sonnet-4.6 [high] | 72.4 | 70.5 |
| 22 | anthropic/claude-sonnet-4.6 [medium] | 71.7 | 69.6 |
| 23 | qwen/qwen3.6-max-preview [enabled] | 71.6 | 88.1 |
| 24 | z-ai/glm-5.3 [high] | 69.3 | 70.4 |
| 25 | anthropic/claude-opus-4.6 [low] | 68.8 | 72.4 |

Notes: 8-needle, full mode, from the site API (184 entries). Ranked here by the 64k-128k token bin because every model has it; the site API gives no ordering and no update date. Overall score averages over whichever lengths each model was tested at, so is not directly comparable. The 2-needle view returned empty at retrieval.
