Because Google models are crap. They are on par with local shit models. Last time I tried they were unable to even hold a longer conversation, because they would enter loop and reiterate last answer regardless of the prompt etc.

Then basic mistakes (inferring wrong information from the given text, hallucinations...). Gemini is okay for some absolute basics and the rest of the work goes to actual frontier models.

How long ago did you try it? Gemini models seem extremely capable to me. I've been using them in ag for over a year and to me the results are great. switching to opus in ag doesn't produce any better results for me