Google’s 6 November 2026 Gemini API notice says requests to gemini-embedding-2-preview now route to gemini-embedding-2. Google asks developers to update the model identifier. It says the API schema, pricing and embedding space are unchanged, so this transition requires no other code changes or vector re-indexing.
The deprecation table confirms the redirect. The stable model launched on 22 April; today’s event is the preview migration.
Keep the change small and observable
For a team maintaining document search, this is a useful housekeeping task. Our suggested approach is to separate the model-name update from unrelated changes to the documents, chunking rules or index configuration. That makes any unexpected result easier to investigate.
- Find the active setting. Check deployment configuration, scheduled ingestion jobs and query services, not just a notebook example. Record which environment uses the preview identifier.
- Update one controlled environment. Use the documented stable name. Confirm the application starts and its request path uses the intended setting before widening the change.
- Try a small retrieval sample. Use non-sensitive test documents and familiar queries with expected matches. Check the application’s error logs and displayed results. Record observations rather than assuming that a successful request proves the whole workflow correct.
- Keep index maintenance separate. Do not schedule a large rebuild solely because this notice uses the word deprecation. If another change requires rebuilding, document that reason independently.
These are proposed deployment checks. We have not run API calls, measured retrieval quality or checked a customer bill. Google’s compatibility statement is specific to this preview-to-stable transition; it is not a general promise about moving between arbitrary embedding models.