Gemini Embedding 2 preview retires without a vector rebuild
Google retired the gemini-embedding-2-preview identifier on November 6, 2026. Requests using it now route to the stable gemini-embedding-2 model. Google says this specific switch preserves the API schema, pricing and embedding space, so existing preview users do not need to rebuild their vector index. Official release notes
The practical action is to replace the preview model string in application configuration. The stable model itself is not a November launch: Google lists its release as April 22, 2026. The new event is the retirement and routing change for the preview endpoint. Model retirement schedule
Which migration avoids rebuilding the index?
For a system already using Gemini Embedding 2 Preview, Google describes a compatible move to the stable identifier. This matters because a vector index stores numerical representations of documents or other content. Recomputing those representations can mean another processing bill and an operational migration.
A different rule applies to gemini-embedding-001. Google says its embedding space is incompatible with Gemini Embedding 2. Moving from that older model requires embedding the existing data again. Vectors from the two models cannot be directly compared. The November 6 notice does not remove that requirement. Embedding migration guide
Start by checking the exact model identifier used for both stored documents and incoming queries. A project described internally as “Gemini embeddings” may contain more than one generation. Keeping that inventory explicit helps avoid applying the preview compatibility promise to the wrong index.
What the stable model supports
Gemini Embedding 2 maps text, images, video, audio and PDFs into a shared space for similarity searches. Google lists an 8,192 token input limit and output dimensions from 128 to 3,072, with 768, 1,536 and 3,072 recommended. These are the existing model’s capabilities, rather than additions announced with the retirement. Model specifications
If an application is also changing from the older 001 model, review request behavior separately. Gemini Embedding 2 uses prompt instructions for text task types instead of the older task_type field. It can also aggregate multiple inputs into one embedding. Google’s guide explains how to request separate embeddings and how query and document formatting should match the retrieval task.
Prices depend on the input
Google’s current standard paid rates are $0.20 for text, $0.45 for images, $6.50 for audio and $12 for video, each per million input tokens. The listed Batch rates are half those amounts. These are USD API input prices, not the full cost of operating a search system. Database storage, indexing and other infrastructure remain separate considerations. Official pricing
Google’s statement that pricing is unchanged concerns the preview to stable transition. It does not mean every input modality costs the same. A workload mixing documents, pictures and recordings needs a budget based on what it actually sends.
A focused deployment check
For preview users, update the model name, keep the existing vector dimensions and preprocessing choices consistent, and run representative queries through the application’s normal retrieval path. Check error handling and usage reporting as part of that rollout. These are practical validation steps, not evidence of a new quality gain or a ByteForward performance test.
Teams also migrating text generation should keep the schedules separate. The Gemini 2.5 Pro and Flash migration guide covers a different December 1 cutover with replacement models and request changes. The embedding preview route has already changed, and its compatibility statement is narrower.
AI-generated editorial illustration by ByteForward. Google and Gemini are trademarks of Google LLC. ByteForward is not affiliated with or endorsed by Google.
