Google has released Gemini Embedding 2 to general availability, marking the production launch of its latest text embedding model. The model, which had been in preview since earlier announcements, is now accessible through Google AI Studio and the Gemini API [source].

Gemini Embedding 2 supports embedding generation for text inputs up to 2,048 tokens in length. Google states the model delivers improved performance on retrieval and semantic similarity tasks compared to its predecessor, though specific benchmark comparisons were not detailed in the announcement. The model outputs 768-dimensional vectors by default.

The release includes support for multiple languages and introduces task-type parameters that allow developers to optimize embeddings for specific use cases including retrieval, classification, and clustering. Google notes the model can handle both short queries and longer document passages within the token limit.

Pricing is set at $0.00001 per 1,000 characters for the standard model variant. Google also announced a reduced-dimension version outputting 256-dimensional vectors at a lower price point for applications where full dimensionality is not required.

The general availability follows a preview period during which developers tested the model in non-production environments. Google has not disclosed the size of the model or specific architectural details beyond confirming it is part of the Gemini model family.

Developers using the preview version will need to update their API calls to reference the GA model identifier. Google states existing preview endpoints will continue functioning but recommends migration to the production version for stability guarantees. No breaking changes to the API surface were announced with the GA release.