What is Gemini Embedding 2?
Gemini Embedding 2 maps text, images, audio, video and PDF pages into one vector space, so a text query can find a photo, a video clip or a spoken passage without a transcript or caption in between. It covers over 100 languages and takes up to 8,192 tokens per request: at most 6 images, 120 seconds of video, 180 seconds of audio or one 6-page PDF.
It is Google's current embedding model and its named replacement for the text-only gemini-embedding-001. The two produce incompatible vectors, so switching means re-embedding everything. Its knowledge stops at November 2025.
What changed from gemini-embedding-001?
It adds images, audio, video and PDFs, takes four times the input, and scores higher, but its vectors do not mix with the old model's.
- Inputs
- The old model read text only; this one also takes images, audio, video and PDFs.
- Input size
- Up to 8,192 tokens per request, against 2,048.
- Scores
- 69.9 against 68.4 on MTEB Multilingual and 84.0 against 76.0 on MTEB Code, in Google's table.
- Migration
- The vector spaces are incompatible, so you re-embed all data, and task_type gives way to an instruction in the input.
- Status
- Google lists Gemini Embedding 2 as the replacement for gemini-embedding-001, which now has a shutdown date.
Who is Gemini Embedding 2 best for?
Gemini Embedding 2 is best for developers building search, RAG or classification over mixed media who want text, images, audio, video and PDFs in one vector space.
Search across media types
A text query can match images, video clips, audio or PDF pages, because all of them land in the same embedding space.
RAG over mixed documents
It embeds PDFs of up to 6 pages and up to 6 images per request directly, with no separate captioning or transcription step.
Multilingual apps
One model covers more than 100 languages, so you can index content in several languages together.
Indexes that must stay small
You can cut vectors from 3,072 dimensions down to 1,536 or 768 to save storage, the sizes Google recommends for quality.
What's included with Gemini Embedding 2?
Gemini Embedding 2 takes five input types, lets you pick the vector size, follows task instructions written into the input, and runs at half price through the Batch API.
- Five input types
- Text, images (PNG, JPEG), audio (MP3, WAV), video (MP4, MOV) and PDFs, alone or mixed in one request.
- Pick the vector size
- Anything from 128 to 3,072 dimensions, and shorter vectors come back already normalized.
- Task instructions
- Write the task, such as code retrieval or search result, into the input text; the old task_type setting is gone.
- One vector or many
- Several parts in one request give one combined embedding; wrap each part on its own to get a vector for each.
- Batch API
- Half-price embedding on the paid tier for jobs that can wait.
- Works with common vector tools
- Google names LangChain, LlamaIndex, Haystack, Weaviate, Qdrant, ChromaDB and its own Vector Search.
How much does Gemini Embedding 2 cost?
Gemini Embedding 2 is free on the Gemini API free tier, and on the paid tier text costs $0.20 per million tokens, or $0.10 through the Batch API.
- Text, image, audio and video input at no cost
- Lower rate limits than the paid tier
- Inputs may be used to improve Google products
- Images $0.225 per million tokens
- Audio $3.25 and video $6.00 per million tokens
- Paid tier only, half the Standard rate
- Images $0.45 per million tokens, about $0.00012 each
- Audio $6.50 per million tokens, about $0.00016 a second
- Video $12.00 per million tokens, about $0.00079 a frame
PDF pages are billed at the image rate. The Gemini API reads video at 1 frame a second for clips up to 32 seconds, so a 30-second clip costs about $0.024 at Standard rates. Paid-tier inputs are not used to improve Google's products. Enterprise use through Gemini Enterprise Agent Platform is priced separately, with volume discounts arranged through Google's sales team. Prices are in US dollars. Full pricing on ai.google.dev. We checked these prices on .
How do readers rate Gemini Embedding 2?
No one has rated Gemini Embedding 2 yet. Be the first.
Sign in to rate. One rating per account.
Gemini Embedding 2 pros and cons
It embeds media directly and leads most of Google's published benchmarks, but it is API only, takes short inputs, and needs a full re-embed to switch from gemini-embedding-001.
Pros
- Embeds images, audio, video and PDFs directly, with no transcript or caption step; Hacker News builders used it to search dashcam footage and photo libraries by plain description.
- Leads Google's published table on text, image, video and speech retrieval, including 84.0 on MTEB Code against 76.0 for gemini-embedding-001; it trails only on ViDoRe v2 document search, 64.9 to Voyage Multimodal 3.5's self-reported 65.5.
- Cheap to start: free on the Gemini API free tier, then $0.20 per million text tokens, and developers on Hacker News put video search at under $3 per hour of footage.
- Shorter vectors keep most of their quality and come back normalized, so one paid call can feed indexes of different sizes.
Cons
- API only, with no open weights: every file goes to Google, and on the free tier Google may use it to improve its products; several Hacker News commenters asked for a local model for that reason.
- Small inputs for a model this broad: 8,192 tokens, 120 seconds of video, 180 seconds of audio or 6 PDF pages per request, so long files must be chunked, and anything over the limit is cut without warning.
- Moving from gemini-embedding-001 means re-embedding all your data, because the two vector spaces do not match, and the task_type setting no longer works.
- During the preview, two testers on Hacker News found its response time too slow for search-as-you-type or live video indexing.
What are the best Gemini Embedding 2 alternatives?
Four tools on Recaply do Gemini Embedding 2's job. EmbeddingGemma 2 is open source, so you can run it on your own server.
CohereFreeChoose it when you embed long text documents: Embed 4 takes up to 128K tokens per input and costs $0.12 per million text tokens, though it handles text, images and PDFs but not audio or video.
Amazon NovaFreemiumChoose it when your stack runs on AWS: Nova Multimodal Embeddings on Amazon Bedrock takes the same five input types, and its async API splits long text, video and audio into segments for you.
PineconeFreeChoose it when you want embedding and storage in one service: its indexes can embed text with hosted models such as llama-text-embed-v2 as you store it, and the Starter plan is free.
EmbeddingGemma 2Open SourceChoose it when you need a 740M embedding model for text, images, audio and video that runs on the device.
Questions about Gemini Embedding 2
What is Gemini Embedding 2?
Gemini Embedding 2 is Google's first natively multimodal embedding model: it turns text, images, audio, video and PDFs into vectors in one shared space, for developers building search, RAG and classification.
Who is Gemini Embedding 2 for?
Gemini Embedding 2 is best for developers building search, RAG or classification over mixed media who want text, images, audio, video and PDFs in one vector space.
Is Gemini Embedding 2 free?
Yes. Gemini Embedding 2 is free on the Gemini API free tier for text, image, audio and video input, with lower rate limits, and Google may use those inputs to improve its products. The paid tier costs $0.20 per million text tokens, or $0.10 with the Batch API.
How much does Gemini Embedding 2 cost?
Gemini Embedding 2's paid plans are Batch at $0.10 per million text tokens and Standard at $0.20 per million text tokens.
Is there an open source alternative to Gemini Embedding 2?
Yes. EmbeddingGemma 2 is an open source tool that does the same job, so you can run your own.
What are the best Gemini Embedding 2 alternatives?
The closest alternatives on Recaply are Cohere, Amazon Nova, Pinecone and EmbeddingGemma 2.
Where can I try Gemini Embedding 2?
On its official website, ai.google.dev.
What are the input limits of Gemini Embedding 2?
Each Gemini Embedding 2 request takes up to 8,192 tokens in total: up to 6 PNG or JPEG images, audio up to 180 seconds, video up to 120 seconds, or one PDF of up to 6 pages. Longer videos have to be split into chunks first.
Does Google use data I send to Gemini Embedding 2?
On the free tier, yes: Google says free-tier content is used to improve its products. On the paid tier it is not. Companies that need extra security and compliance controls can use the model through Gemini Enterprise Agent Platform instead.
Can I use Gemini Embedding 2 on Google Cloud?
Yes. Besides the Gemini API, it is generally available on Gemini Enterprise Agent Platform as gemini-embedding-2, in the global, US and EU locations. There it also supports batch jobs and can pull the audio track out of a video to embed alongside the frames.
Our verdict
Gemini Embedding 2 is the easy pick for search or RAG over mixed media, and the default for new Gemini projects.
It leads most rows of Google's published retrieval table, handles audio and video without transcripts, and costs $0.20 per million text tokens after a free tier.
- Skip it for
- Anyone who must keep data on their own hardware, or who needs long text documents embedded in one piece.
- Check first
- If you run gemini-embedding-001 today, budget for re-embedding everything before you switch.
Cohere Embed 5
OpenAI Privacy Filter
Cua Spaces
Spacefast
Gemini Computer Use
Gemini CLI