hey y'all, question for the context engineers here : have new s-o-t-a models, like Opus 4.6/Kimi 2.5/Minimax 2.5/GPT 5.3, higher context efficiency rates or is the main "goldy locks" zone still around 30-40% regardless of the models ?