Context Arena MRCR (2-needle): leaderboard
OpenAI MRCR benchmark: finds and distinguishes between 2 identical pieces of information in long conversations (8K–1M tokens). Tests multi-round co-reference resolution via AUC across context lengths.
Metric: AUC@1M (%). Source: contextarena.ai. Saturation forecast: Around January 2027. 89 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) (High) | 93.8 | #82 (Gemini 3 Flash (Preview)) |
| 2 | Gemini 3 Flash (Preview) (Medium) | 90 | #82 (Gemini 3 Flash (Preview)) |
| 3 | Gemini 3 Flash (Preview) (Low) | 89 | #82 (Gemini 3 Flash (Preview)) |
| 4 | Gemini 3.1 Pro (Preview) (High) | 87.8 | #58 (Gemini 3.1 Pro (Preview)) |
| 5 | Gemini 3.1 Pro (Preview) (Medium) | 87.7 | #58 (Gemini 3.1 Pro (Preview)) |
| 6 | Gemini 3.1 Pro (Preview) (Low) | 87.4 | #58 (Gemini 3.1 Pro (Preview)) |
| 7 | Gemini 3 Pro (Preview) | 85.5 | #66 |
| 8 | Gemini 3 Flash (Preview) | 85.5 | #82 |
| 9 | Gemini 2.5 Flash (Preview 05-20) | 78.3 | #216 |
| 10 | Gemini 2.5 Pro (Preview 06-05) | 77.5 | #106 |
| 11 | Nova 2.0 Lite (Non-reasoning) | 63.5 | #417 (Nova 2.0 Lite) |
| 12 | GPT-4.1 | 53.2 | #248 |
| 13 | GPT-4.1 Mini | 43.6 | #355 |
| 14 | Grok 4.1 Fast | 40.3 | #217 |
| 15 | Llama 4 Maverick | 39.9 | #464 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=context-arena-mrcr-2-needle · How It Works · Data refreshed daily, snapshot 2026-10-20.