Google Gemini API Deprecation Calendar
Google's dates carry one caveat that changes how you plan: most are marked "earliest possible date", not a hard shutdown. Treat them as floors - the model will not disappear before the date, but may run past it. That means you should be migrated by the date without treating it as the last safe day. The list is short compared to OpenAI or the cloud platforms, but the Gemini 2.5 retirement is a big one because so much production traffic still sits on it.
Scheduled retirements
| Date | What retires | Do this before the date | Src |
|---|---|---|---|
| Aug 17 (earliest) | Imagen 4 family (imagen-4.0-*) and Gemini 3 image models | Move image generation to Nano Banana 2 or gemini-3.1-flash-lite-image. Re-test outputs - image model swaps change style, aspect handling and safety filtering. Because this is an "earliest" date, watch the changelog for the confirmed cutoff. | ✓ |
| Oct 16 (earliest) | gemini-2.5-pro, gemini-2.5-flash | The big one. Migrate to gemini-3.1-pro-preview / gemini-3.5-flash. Note the target for Pro is still a preview alias - pin a stable snapshot as soon as one ships, or you inherit a second migration. Re-run evals; the 2.5 to 3.x jump changes behavior and pricing. | ✓ |
| May 7, 2027 | gemini-3.1-flash-lite-preview | Move to the stable gemini-3.1-flash-lite. This is a preview-to-GA graduation, so behavior should match - but confirm the GA snapshot id and drop the -preview suffix. | ✓ |
Known gap: Vertex AI
These dates come from the Gemini API (ai.google.dev). Google's Vertex AI platform publishes no model-retirement table for partner models, so Vertex-specific shutdown dates are unverified here. If you run Gemini through Vertex, the Gemini API dates above are your best signal, but confirm in your Vertex console - the two surfaces version independently. This calendar is maintenance guidance, not legal advice.