In the CorpusQA evaluation, Gemini 2.5 Pro achieved an overall accuracy of 80.19% at a 128K context length and 50.73% at a 1M context length, a drop of 29.46 percentage points.