Nothing too puzzling really. It was a good assistant specialized for a relatively narrow set of tasks and tooling on a much more modest chatbot architecture. Now it has a wider set of tools it can use, but is also computationally more expensive, so it runs slower. Makes it better if you want to do any task, but if you want to do a task like set the timer, the old model is obviously faster

The bigger issue is that it doesn't "know" what its capabilities are, and if you ask it to do something it can't do, it's likely to LARP doing them instead of telling you it can't.

The puzzling part is that it can't fall back to a dumb assistant for stuff it could always do. The slow part has not just been interpretation of the task, Gemini is actually slower at doing the task.