Google says Vertex AI SDK Gemini modules are being removed June 24, 2026. Migrate Gemini apps to the Google Gen AI SDK now. Read the guide.
What is changing
Google is removing the Gemini modules from the Vertex AI SDK on June 24, 2026. If your application still imports Gemini through that path, those imports and helpers will stop working when the modules go away. This is not a soft deprecation you can ignore indefinitely; it is a hard removal date you should plan against now.
The replacement path is the Google Gen AI SDK. The intent is a single client surface for Gemini rather than Gemini-specific modules nested inside the Vertex AI SDK. For teams, the practical effect is a package and import change, plus a review of how models, auth, and generation options are configured.
Who needs to act
You need to migrate if any service, job, or notebook still calls Gemini through the Vertex AI SDK Gemini modules. That includes production APIs, background workers, evaluation scripts, and internal tools that share the same client code. If Gemini is only used through the Google Gen AI SDK already, this removal may not touch you—but confirm with a repo-wide search rather than assuming.
Shared libraries are easy to miss. One team can migrate its service while another still depends on a wrapper that re-exports the old modules. Treat libraries, sample code, and CI fixtures as first-class migration targets, not afterthoughts.
How to migrate with less risk
Start by inventorying every place Gemini is reached through the Vertex AI SDK. Search for the old module paths, client constructors, and generation helpers. List each call site with its runtime (service, job, notebook) and who owns it. That list becomes the migration backlog and the checklist for go-live.
- Swap imports and client setup to the Google Gen AI SDK while keeping model choice and prompts stable at first.
- Re-check auth and project or location configuration so the new client resolves the same environment as production today.
- Compare outputs on a fixed set of prompts before and after the switch so regressions show up as behavior changes, not vague “it feels different” reports.
- Ship behind a feature flag or dual-run path where you can, then remove the old modules only after traffic is clean.
Prefer small, reversible steps over a big-bang rewrite. Migrate one service or library boundary at a time, keep the old path available until the new path is proven, and only then delete the Vertex AI SDK Gemini usage. That reduces the chance that June 24, 2026 becomes an outage date instead of a non-event.
What to finish before the deadline
Work backward from June 24, 2026. Leave time for dependency updates, staging validation, and a production rollout with monitoring on error rates, latency, and empty or malformed responses. Freeze new features that still depend on the old modules so the migration does not compete with unrelated changes.
After cutover, remove dead imports and outdated internal docs so the next engineer does not reintroduce the old path. Treat the Google Gen AI SDK as the only supported way to call Gemini going forward, and keep a short runbook for how to create the client, set credentials, and run a smoke prompt in each environment. Migrating now is cheaper than discovering broken imports on the removal date.