Comment by anana_

The benchmarks they chose are rather cherry picked to not include long context or difficult ones that involve long horizon work or many agent turns, as I suspect this is where the model shows more differences compared to the full fat one