"Then a month or so ago it turned back into being garbage. I waited for it to be fixed but it has remained bad."

You have not provided any evidence for that or explained what you mean. Are you using a specific version of Gemini via a service, api, Gemini app? AI studio? etc? I have noticed zero issues.

But in the last month Gemini 3 flash was released, are you getting confused between models? (Although in my experience Gemini 3 flash is excellent)

The Gemini 3 series are far better than 2.5, which explains the first half

I haven't experienced any degradation myself, but many times people attribute this to companies running quantified versions of the model during peak demand

I haven't noticed it being awful again. Can you give examples?

Gemini has been fantastic for me the last month or so.

Quantizing To decrease cost