Google publishes a discontinuation date for each Gemini version on Vertex AI. Versioned ids retire on schedule while the unversioned alias moves forward, so code pinned to a version breaks and code pinned to an alias changes behavior instead.
We track 84 Google Gemini and Vertex AI models, 50 of them with a published retirement date. 7 need attention now.
Deprecated models still answer but are superseded. Anything inside the 90-day window has a date attached and will stop answering on it.
| Model | State | Retirement | Replacement |
|---|---|---|---|
| gemini-2.5-flash-image | retiring | 2026-10-02 · past due | google/gemini-3.1-flash-image-preview |
| gemini-2.5-flash | retiring | 2026-10-20 · in 12 days | — |
| gemini-2.5-flash-lite | retiring | 2026-10-20 · in 12 days | — |
| gemini-2.5-pro | retiring | 2026-10-20 · in 12 days | — |
| veo-3.1-fast-generate-preview | retiring | 2026-10-22 · in 14 days | google/gemini-omni-1.1-flash |
| veo-3.1-generate-preview | retiring | 2026-10-22 · in 14 days | google/gemini-omni-1.1-flash |
| veo-3.1-lite-generate-preview | retiring | 2026-10-22 · in 14 days | google/gemini-omni-1.1-flash |
These ids no longer answer. If one appears in your code or in something an assistant wrote for you, it is already broken.
Active, announced, or legacy with no imminent date. Legacy models are superseded but still served.
A list of dates only helps once you know which of these models your own traffic hits. One line of code meters the models your app actually calls, then reports the monthly spend sitting on each retiring id and what the move to its replacement costs.
Source: Google — Gemini API deprecations. Every record links the page it was parsed from and keeps Google’s own wording, so a normalization decision stays checkable.