Every SOTA model I've used at launch uses deeper, longer inference then gradually turns down over time, until the next model comes out which seems to be trained on some new data, but mostly performance due to deeper longer inference for another period.