I think a lot is memory really. For fun you can try pointing just about any model against the older kaggle challenges and it’ll come up with more or less the winning solution from memory. It’s here where you see a Gemma 27b perform about the same as opus 5.5