Executive summary

Google has released Gemini Embedding 2 into public preview, a purpose-built embeddings model that maps text, images, native audio, video and document content into a single vector space intended to simplify enterprise retrieval and downstream multimodal workflows. This model is designed as a native multimodal embedder rather than a text-first retrofit: it can consume audio and video natively (without mandatory transcription) alongside images and documents, enabling direct cross-modal queries and reducing error introduced by separate transcription pipelines.

Technical profile

Gemini Embedding 2 emits a 3,072‑dimensional vector by default and implements a Matryoshka-style layered representation that lets teams truncate to smaller footprints (for example 768 or 1,536 dims) to save storage and index costs with modest accuracy tradeoffs. Per-request ceilings are explicit: up to 8,192 text tokens, up to 6 images, up to 128 seconds of video, up to 80 seconds of audio, or a 6‑page PDF; larger assets require chunking and re-indexing. Google positions the model around cross-modal semantic alignment rather than stitched transcription flows, which reduces translation-related retrieval errors and can lower end-to-end latency in pipeline implementations.

Deployment, pricing and access

The preview includes a free, rate-limited experimental tier (roughly 60 requests per minute) plus paid usage tiers: approximately $0.25 per 1M tokens for most inputs and $0.50 per 1M tokens for native audio. Distribution is being handled through Google’s public APIs and Vertex AI channels; however, Google is staging feature exposure — consumer and enterprise access paths differ (AI Pro and Gemini Alpha programs) and some enterprise features require tenant admin opt‑ins. Separately reported product updates across the Gemini family indicate Google is keeping preview rate cards broadly consistent while gating the deepest capabilities behind subscription and staged enterprise programs.

Ecosystem and integrations

Google has integrated the model into developer tooling and enterprise gateways and early integrations with LangChain, LlamaIndex, Weaviate, Qdrant and others reduce adoption friction. Embedding-native search can be combined with Workspace and Drive integrations in the broader Gemini stack to let apps pull contextual signals from Drive, Gmail and Chat — turning passive file stores into active retrieval-and-synthesis layers. Gemini’s multimodal and reasoning improvements (reported in companion updates such as Gemini 3.1 Pro) further increase the utility of embeddings for building synthesis-centric workflows, lowering token usage and iteration costs in some partner pilots.

Enterprise implications and governance

Early adopters report practical gains — for example, some customers cite latency reductions up to 70% and Everlaw reports a recall lift near 20% for legal discovery — but these gains come with migration costs: corpora must be re‑embedded and assets chunked to fit per-request limits. Google emphasizes enterprise controls (data isolation, tenant-admin opt-ins) and multimedia provenance tools such as SynthID watermarking to address compliance. That said, public reporting also highlights rapid, large-scale pilots (a reported unclassified Department of Defense pilot showing ~1.2M distinct users issuing ~40M prompts) that underline governance gaps when adoption outpaces accredited training and policy processes.

Bottom line

For architecture and procurement teams the trade-offs are clear: the model can materially simplify pipelines and enable richer cross-modal search, but realizing those benefits requires one-time reindexing cost, careful truncation and storage planning, and attention to governance and portability. The wider Gemini product updates — reasoning improvements, Workspace integrations, and staged commercial gating — mean the embedding is being launched as part of a broader platform strategy rather than as an isolated model update.