What it is
Gemini Embedding 2 is Google's first multimodal embedding model. It launched in preview on March 10, 2026 and became generally available on April 22 under the ID gemini-embedding-2; the old gemini-embedding-2-preview page remains online. Embedding models turn content into vectors so software can find related material for search, recommendations and retrieval-augmented generation (RAG).
It takes text, images, video, audio and PDF files and places them in one shared vector space, so a text query can find a matching image or video clip. Each request is capped at 8,192 tokens across all inputs, and output can be set anywhere from 128 to 3,072 dimensions, with shorter vectors normalised automatically.
Pricing per million tokens is $0.20 for text, $0.45 for images, $6.50 for audio and $12 for video, with Batch at half price. It is on the Gemini API free tier, where Google may use inputs to improve its products.
Our review
Gemini Embedding 2 earns its place on mixed media, not on text. Text, images, audio, video and PDFs land in one space, which turns 'find the moment in this recording where the speaker mentions refunds' into a single nearest-neighbour lookup instead of a transcription pipeline. It has been generally available since April 22, so the preview caveat that hung over it through the spring no longer applies.
For text alone it is the expensive option. At $0.20 per million tokens it costs more than voyage-4-large at $0.12, and Google's docs publish no MTEB score for it at all; the only MTEB table on the embeddings page is for the older gemini-embedding-001. Vectors from 001 are not compatible, so moving up means re-embedding the whole corpus.
The per-request limits shape the architecture more than the price does. Each call takes at most six images, 180 seconds of audio, 120 seconds of video sampled to 32 frames, or one PDF of up to six pages, and anything over 8,192 tokens is silently truncated. Video files also lose their soundtrack: the audio track is ignored, so a meeting recording needs embedding twice, once as video and once as audio, to be searchable by both.
Media pricing is lower than the per-token rate suggests. Google converts video to about $0.00079 a frame, so a two-minute clip at 32 frames costs about 2.5 cents, and Batch halves that. The cost risk is not one clip but re-indexing an archive every time you change chunk size.
Where it earns its place
A media-archive engineer at a regional broadcaster with 10,000 hours of footage
Let producers find a shot by describing it in a sentence
Chunk footage into overlapping two-minute segments and embed each at 32 frames. At the published per-frame rate that is roughly $7,600 for the archive, or about $3,800 through Batch, plus a separate audio pass if speech must be searchable.
A product manager at a 30-person e-commerce marketplace
Add 'search by photo' alongside text search for 2 million listings
Embed each listing's main image (about $0.00012 per image, $0.00006 in Batch) and title into the same space, so a shopper's photo and a typed query hit one index. Two million images cost about $240 at standard rates.
A developer at an EU legal-tech startup
Embed scanned contracts with tables and stamps for retrieval
Send PDFs page by page, which Google says gives the best quality, with OCR always on. EEA-facing apps must use the paid tier, where Google does not use inputs to improve its products.
Before you commit
- Setup
- One call to embedContent with model gemini-embedding-2 in the Gemini SDKs; the task_type parameter is not supported, so text tasks use prefixes such as 'task: search result | query:'.
- Integrations
- Gemini API, Google AI Studio, the Batch API and Gemini File Search, which uses this model for multimodal search with page-number and media citations.
- Data and privacy
- On the paid tier Google does not use prompts or responses to improve products but logs them for a limited, unstated period for abuse monitoring. Free-tier inputs may be used to improve Google's products and machine-learning technology and read by human reviewers, and apps serving users in the EEA, Switzerland or the UK must use paid services.
- Cost at scale
- A 100M-token text corpus costs $20 to embed ($10 in Batch); 10,000 hours of audio costs about $5,760 at $0.00016 a second; video runs about $0.00079 a frame. Re-embedding on every chunking change is the cost to plan for.
- Limits
- 8,192 tokens per request across all inputs, silently truncated; six images, 180 seconds of audio, 120 seconds of video (32 frames, audio ignored) or one six-page PDF per request. No published MTEB score.
What's good
- Text, image, video, audio and PDF in one vector space, generally available
- 128 to 3,072 dimensions with automatic normalisation
- Batch at half price; video at about $0.0008 a frame
What's not
- No published retrieval benchmarks
- 8,192-token cap per request, silently truncated; video audio tracks ignored
- Pricier than voyage-4-large for text; incompatible with gemini-embedding-001 vectors
Best for
Developers building search or recommendations across mixed media such as images, recordings and video.
Alternatives
Gemini Embedding 2: common questions
- Is Gemini Embedding 2 worth it?
- Gemini Embedding 2 is the simplest way to search video and audio with a sentence, and an overpriced way to search plain text. We rate it 4.3 out of 5.
- How much does Gemini Embedding 2 cost?
- Per 1M tokens: text $0.20, image $0.45, audio $6.50, video $12; Batch half price; free tier with limits (inputs used for product improvement).
- What are the downsides of Gemini Embedding 2?
- No published retrieval benchmarks; 8,192-token cap per request, silently truncated; video audio tracks ignored; pricier than voyage-4-large for text; incompatible with gemini-embedding-001 vectors.
- Who is Gemini Embedding 2 best for?
- Developers building search or recommendations across mixed media such as images, recordings and video.
- How much does Gemini Embedding 2 cost for a team or in production?
- A 100M-token text corpus costs $20 to embed ($10 in Batch); 10,000 hours of audio costs about $5,760 at $0.00016 a second; video runs about $0.00079 a frame. Re-embedding on every chunking change is the cost to plan for.
- Is Gemini Embedding 2 safe for company data?
- On the paid tier Google does not use prompts or responses to improve products but logs them for a limited, unstated period for abuse monitoring. Free-tier inputs may be used to improve Google's products and machine-learning technology and read by human reviewers, and apps serving users in the EEA, Switzerland or the UK must use paid services.
- What are the best Gemini Embedding 2 alternatives?
- The closest alternatives we have reviewed are Cohere Embed 5, voyage-4-large, text-embedding-3-large and EmbeddingGemma 2.
Answers checked Oct 22, 2026.
Make Gemini Embedding 2? Show this rating on your site
