A model you picked yesterday isn't in today's list. On September 18, a thread appeared on the AI Studio community forum saying exactly that: Gemini 2.5 Pro was there in the afternoon and gone by evening. Replies mentioned conversations freezing when the picker emptied. Reading it, I felt that small sinking feeling in my stomach.
What I'd like to say first is that "not in the list" and "deprecated" are different things. Here is how I tell them apart, and the order in which I now keep my history safe.
Check the official page before deciding
I opened the deprecations page on October 21, 2026. It listed these rows:
| Model | Release date | Shutdown date | Recommended replacement |
|---|---|---|---|
| gemini-2.5-pro | June 17, 2025 | None announced | Blank |
| gemini-2.5-flash | June 17, 2025 | None announced | Blank |
| gemini-2.5-flash-lite | July 22, 2025 | None announced | Blank |
The same page says access to the 2.5 models is limited to people who have actively used them in the past, that the models are not deprecated, and that they'll keep being served until further notice. For new projects it points to 3.5 Flash-Lite or 3.8 Flash.
So no shutdown date exists, but visibility has been narrowed. New users, or accounts judged inactive, may simply not see the models.
Which one are you seeing?
Checking both the UI and the API speeds this up.
- In AI Studio, click the blank area of the model picker to expand the list. A forum user reported that this worked; I haven't reproduced it myself.
- Call the same model name with a key from a different project and compare results.
- If you get a 404, suspect the access restriction before a typo. I wrote up that triage in the 2.5 Flash 404 with no announced shutdown date.
Some forum posters guessed it was about plan tiers or cost. That rested on a video, not on anything official, so I'm leaving it out of my assumptions.
Keep history independent of the model
As an indie developer, I tend to leave prototype conversations sitting in AI Studio. I knew a model could disappear from under them, and I hadn't prepared. That did not go well. Now I do three things regardless of what any model does:
- Export conversations I care about to a document at milestones. Check who can see the export first, as I describe in checking sharing before sending an export.
- Write the model name and date on the first line of each conversation, so I know what to paste where later.
- For API work, pin the model name and log the model that comes back. I covered that in pinning defaults and detecting silent swaps.
Models may move; the record stays on my side. I try to hold that line even in busy weeks.
One next step
Pick a single conversation you can't afford to lose, export it, and add the model name and date at the top. Official pages change, and this reflects October 21, 2026, so open the deprecations page first whenever something stops working.