Small rant in a jovial way but…
Since I started my project gemini-1.5 has been deprecated and so I switched to gemini-2.5-flash-lite for my application to use. I am in a constant battle with the AI to stop changing the model in my code. Nothing seems to stop it from doing so, I have created very clear gemini.md file that tells it what model I am using but it still does it due to the fact that after you are about 3-4 edits in, the ai’s memory gets very wonky and it no longer remembers the rules. Then I remind it that 1.5 is deprecated and then it argues with me that it does in fact exist but it does not. Sometimes to be an ass I let it make the change and show it the logs of the failure where it says 1.5 isn’t available and it apologizes and tries to correct it AGAIN to 1.5 or something generic like gemini-flash or gemini-pro.
As frustrating as it is, I do chuckle at the non-sense that Gemini does not know what Gemini models it has available.
That said, this is an annoyance that I am sure others run into, and it can be quite jarring after a small edit or change and the model number is changed and your latest changes do not work. You trouble shoot it and the code assist is struggling to find the reason its not working when in fact it is the model that it thinks still exists 
,
It would just be very helpful if you kept the model in the know about what is actually available, I mean hell, if it does not know that 1.5 is deprecated, how will it know all the new functions and updates to next.js or react etc. to help you fix the code you have written when needed… if we are using this code assist to help and assist us, we should have a model that actually knows real time events. It does not appear to be grounded in the slightest and is working off very old information. Based on my conversations with the code assist I have been able to deduce that vast majority of its working knowledge stops 1st or 2nd quarter of 2024 which is almost 2 years out of date. There have been large advancements in AI and code since that time with a lot of major releases. Hell, it didn’t even know about the React & Next.js Vulnerability (CVE-2025-55182) issue that recently popped up. When I asked the AI to analyze my changes to make sure I was in compliance, it called the CVE-2025-55182 issue “fictitious” because it was marked as a “future date”.
It would be very helpful if the model we use in the built-in setting was grounded or at very least set to no more than a month or so out from the current date.
I get AI isn’t perfect and I use it as more of a guardrail than anything and a QC agent, it still is frustrating when it changes things based on outdated information and tries to use depreciated models etc.
