I did not realize that AI Studio was an option, so until now I have been utilizing 2.5 Pro through Gemini Advanced at gemini.google.com/app. I noticed the tool has become dramatically worse since it’s switched 2.5 Pro from experimental to preview mode.
I am noticing these issues to the point where it is almost unusable now. It also seems like the model’s ability to generate functional code has severely decreased. This did not start happening to me until yesterday, so I don’t know if the updates were rolled out in a phased schema.
Now none of the code it’s outputting is usable; initially, it was outputting code that was essentially bugless, but not anymore. I was trying to get it to put together a simple Dockerfile to set up a Tailscale connection and test TCP over it, but the new version could not get this right to save its life.
I also noticed issues of it hallucinating previous code across context windows that I did not provide to it, these just started as well. It was insistent that my previous code had an API key for an email service that I never heard of and that I gave it a version that contained this key and it definitely did not.
I also noticed that it seems to no longer completely view codebases that I provide to it. I will ask it information about a file and it will give a hallucinated answer instead of being based on the actual file content.
