Like I mentioned, we omitted routing from the benchmark and evaluated with our strongest model modality. Also, randomly selecting between Fable and Gemini 3.7 Flash wouldn't preserve quality.