In the same week, Gemini changed both how long the model thinks and where you can call it from.
Google DeepMind introduced Gemini 3.8 Live alongside 3.8 Live Extended Thinking, explicitly splitting a fast-response tier from one that allocates longer deliberation. In the same week, a Windows build of Gemini shipped, letting users invoke it directly from whatever they are working on rather than opening a browser first.
These land on different layers. Live and Extended Thinking widen the choice of how much the model thinks; the Windows build changes where the model gets called from. Either one alone moves usage less than it looks. A choice of thinking time does not help if invoking it is a chore, and easy invocation does not stick if the response does not fit the job. Note: which tier the Windows build defaults to has not been stated.
Which workloads make Extended Thinking the default. Whether the Windows invocation path extends to other operating systems and to existing applications.