Even models with million-token context windows like Gemini 2.5 face attention dilution in long-running complex tasks — as the context continues to grow, the model's attention to early information drops significantly.