← All benchmarks

MRCR v2 (8-needle, Context Arena) - 128K AUC

General QA

Context Arena's own runs of Google DeepMind's open-sourced MRCR v2 8-needle task, scored by MRCR string similarity and summarized as the area under the score-versus-context curve from 8K to 128K tokens: trapezoids on a linear token axis over its 8K, 16K, 32K, 64K and 128K bins, 484 tests in all. Google's model cards quote this value as the 128K cumulative score for non-Gemini models, but these evaluator runs use Context Arena's own bins and reasoning settings, so they form a factor group apart from vendor-reported MRCR v2.

Top models (higher is better)