Google's November 2, 2026 Gemini API release notes deprecate Gemini 2.5 Pro and Flash and say both are no longer selectable in AI Studio. From December 1, requests to gemini-2.5-pro will route to gemini-3.8-flash; gemini-2.5-flash will route to gemini-3.5-flash-lite, at the replacement models' rates.
Why a working request can still change your service
Our assessment: administrators of document-processing, support and other AI integrations should treat this as a controlled service change. A successful HTTP response is only the first check. A replacement model can produce different answers, take a different amount of time or change the cost of a completed job. Do not use an unchanged model name as proof that the underlying behavior is unchanged.
Check compatibility before the switch
Google's 3.8 Flash migration guide calls for replacing thinking_budget with thinking_level, reviewing function calls and preserving thought signatures. Its checklist also removes sampling parameters and prefilled model turns.
The 3.5 Flash-Lite guide likewise covers parameter and conversation changes. It distinguishes ignored, deprecated sampling settings from unsupported prefilled model turns that return an error. Check the target's rules rather than assuming every old field fails identically.
A practical migration checklist
- Inventory model names in applications, scheduled jobs and vendor-managed integrations.
- Ask the responsible team to test representative, non-sensitive inputs, including Hebrew and mixed-language documents where relevant.
- Compare answer correctness, output format, tool permissions, latency and cost per completed task.
- Choose an explicit replacement after testing, with monitoring and a fallback that does not depend on the retiring model.
We have not tested these migrations. December routing is announced, not yet observed. Scope: Gemini API, not consumer subscriptions or every Google hosting service.