MRCR
Can the model retain and reason over evidence spread across a very long input?
Long Context15 ranked models21 reported values4 reportsHigher is better
Top rankings
One row per model, using its best reported score across effort settings.
| Rank | Model | Best score | Best reported setting |
|---|---|---|---|
| 1 | Gemini 3.7 FlashGoogle | 97.0% | Reported configurationGoogle report1 value · 1 reportAug 13, 2026 · source |
| 2 | GPT-5.5OpenAI | 94.8% | Reported configurationGoogle report1 value · 1 reportMay 19, 2026 · source |
| 3 | GPT-5.6 TerraOpenAI | 93.5% | Reported configurationGoogle report1 value · 1 reportAug 13, 2026 · source |
| 4 | Gemini 3.6 FlashGoogle | 91.8% | Reported configurationGoogle report2 values · 2 reportsAug 13, 2026 · source |
| 5 | Gemini 3.1 ProGoogle | 84.9% | Reported configurationGoogle report3 values · 3 reportsJul 21, 2026 · source |
| 6 | Claude Sonnet 4.6Anthropic | 84.9% | Reported configurationGoogle report2 values · 2 reportsMay 19, 2026 · source |
| 7 | Claude Opus 4.6Anthropic | 84.0% | max effortGoogle report1 value · 1 reportFeb 19, 2026 · source |
| 8 | GPT-5.2OpenAI | 83.8% | xhigh effortGoogle report1 value · 1 reportFeb 19, 2026 · source |
| 9 | Claude Sonnet 5Anthropic | 81.5% | Reported configurationGoogle report2 values · 2 reportsAug 13, 2026 · source |
| 10 | Grok 4.5SpaceXAI | 81.4% | Reported configurationGoogle report1 value · 1 reportJul 21, 2026 · source |
| 11 | Gemini 3.5 FlashGoogle | 77.3% | Reported configurationGoogle report2 values · 2 reportsJul 21, 2026 · source |
| 12 | Gemini 3 ProGoogle | 77.0% | high effortGoogle report1 value · 1 reportFeb 19, 2026 · source |
Effort curve
Every sourced cost-linked effort value for this exact version. Lines connect complete sweeps only.
No cost-linked effort sweep for this version.
Definition and comparison boundarypublic methodology
Can the model retain and reason over evidence spread across a very long input?
Long-context retrieval and reasoning, not context-window size alone. Higher is better. The value is the percentage reported in this lab's table.
Harness, tools, prompts, attempt count, and benchmark version can materially change the result. Compare within one reporting context.
Google-computed cumulative 128K score.
Method / source