The problem is Google is treating all those functions ("tool calling") as just any other response, they didn't even bother training the model to recognize and process the requests properly using the tools. They did the Google thing of just putting the function in and calling a day and forgetting about it.

So without proper reinforcement learning it responds with the same casual randomized answer-shaped sequence of words as ever, you're just lucky if the part of it that knows to look up the tool call instructions responds, and luckier if it does it right AND submit the tool call formatted right. If it forgets to look up it's tool call instructions, it might randomly tell you it can't do it because you got an answer from a part of the network trained on data from when there weren't models that could.