Embedding Models: How to Choose One

Definition: An embedding model is the neural network that turns text, images, or other content into vector embeddings, lists of numbers where similar meanings land close together. Embeddings are points in a latent space. Choosing an embedding model means trading off quality, vector size, input length, languages, license, and price per million tokens.

TL;DR: For most teams, the embedding model matters less than the choice to test it on your own data. Prices we checked on 2026-10-19 range from $0.02 to $0.20 per million input tokens for hosted text models. Shrinking Gemini's embeddings from 1,536 to 256 numbers cost about 2 MTEB points. Start with what a vector embedding is if the term is new.

Think of hiring a librarian. Every librarian can shelve books, but one knows medicine, one reads six languages, and one is fast but only skims titles. The shelves you end up with depend on who you hired, and once the shelving is done, switching librarians means reshelving everything. An embedding model is that librarian, and the "reshelving" is re-embedding all of your content.

Why Choosing an Embedding Model Matters in 2026

The choice matters because it is hard to undo. Every document in your search index is embedded by one model, and a query must be embedded by the same model to match. Change models and you re-embed the whole collection. The model also sets your storage bill, your search speed, how long a passage it can read at once, and which languages it understands.

The market in 2026 is crowded and moves fast. OpenAI, Google, Cohere, and Voyage AI each ship hosted models, and open-weight models such as BGE-M3, Qwen3-Embedding, and nomic-embed-text run on your own hardware. The MMTEB benchmark (February 2025) found that the best publicly available model on its multilingual tasks was multilingual-e5-large-instruct with only 560 million parameters, ahead of far larger models. Bigger is not automatically better here.

The Five Numbers That Decide the Choice

Most of the decision comes down to five numbers you can read off a model card.

A sixth factor is not a number: the license, if you plan to run an open-weight model yourself.

Hosted Embedding Models Compared (Checked 2026-10-19)

We checked each vendor's own documentation and pricing page on 2026-10-19. Prices are for text input, standard tier, unless stated. Vendors change prices, so read the linked page before you commit.

Notes from the vendor pages:

• OpenAI lists MTEB averages of 62.3% for text-embedding-3-small and 64.6% for text-embedding-3-large, and estimates about 62,500 and 9,615 pages per dollar at roughly 800 tokens a page. Prices are on its pricing page.
• Google says gemini-embedding-2 accepts text, images, video, audio, and PDFs, and the pricing page shows a free tier on the standard rate. The older gemini-embedding-001 is text only, with a 2,048-token limit.
• Voyage AI gives 200 million free tokens on its current models, per its pricing page, updated 2026-09-29. It also lists an open-weight voyage-4-nano.
• Cohere lists embed-v5.0-pro, embed-v5.0-fast, and embed-v4.0, all reading text and images. The pricing page showed per-instance Model Vault rates and a "Speak to sales" link, not a per-token embedding price, when we read it.

Open-Weight Embedding Models

Open-weight models run on your own servers. You pay for hardware instead of tokens, and your text never leaves your systems.

Open-weight leaderboard positions change often. The June 2025 rank for Qwen3-Embedding is a snapshot from its own model card, not a current standing.

Dimensions: Bigger Vectors, Bigger Bills

More dimensions can hold more detail, but every extra number costs storage and search time. Many 2026 models let you pick a smaller size from the same model, thanks to Matryoshka Representation Learning. Aditya Kusupati and colleagues introduced it in Matryoshka Representation Learning (May 2022): a model is trained so that the first few numbers of each embedding already carry most of the meaning, like nested dolls. On ImageNet-1K they reported embeddings up to 14 times smaller at the same accuracy.

Google's documentation shows what the trade looks like for gemini-embedding-001 on MTEB:

From 1,536 down to 768 dimensions, the score barely moves (68.17 to 67.99). Below 512, it starts to fall faster. Nomic's model card shows the same shape: 62.28 at 768 dimensions and 56.10 at 64. The storage side of the trade is plain arithmetic:

Storage for 1 million chunks, float32 (4 bytes per number)

  3,072 dims  x 4 bytes x 1,000,000  =  about 12.3 GB
  1,536 dims  x 4 bytes x 1,000,000  =  about  6.1 GB
    768 dims  x 4 bytes x 1,000,000  =  about  3.1 GB
    256 dims  x 4 bytes x 1,000,000  =  about  1.0 GB

  Cost to embed 10 million tokens once (vendor list prices, 2026-10-19)

  text-embedding-3-small   $0.02 / 1M  ->  $0.20
  voyage-4                 $0.06 / 1M  ->  $0.60
  text-embedding-3-large   $0.13 / 1M  ->  $1.30
  gemini-embedding-2       $0.20 / 1M  ->  $2.00  (batch: $1.00)

For most document collections, the embedding bill is small next to the storage and search bill that follows. Pick the smallest size that passes your own quality test.

Benchmarks: What MTEB Tells You and What It Does Not

MTEB is the standard yardstick. Niklas Muennighoff, Nouamane Tazi, Loïc Magne, and Nils Reimers introduced MTEB: Massive Text Embedding Benchmark in October 2022 with 8 task types, 58 datasets, and 112 languages. Their main finding still holds: no single embedding method wins on every task. The 2025 MMTEB expansion grew the benchmark to over 500 tasks across 250+ languages. Current scores live on the MTEB leaderboard.

Use the benchmark to build a shortlist, then stop trusting it. A model's average hides large swings between tasks, a retrieval score says little about clustering, and public benchmarks rarely look like your contracts, tickets, or product catalog. Scores reported by a vendor about its own model are a starting point, not an independent result.

How to Choose an Embedding Model

A short checklist covers most cases:

1. Name the content. Language mix, document length, and whether images or PDFs matter decide your shortlist.
2. Check the input limit. If your passages are long, a model with an 8,192-token limit forces smaller chunks than one with 32,000.
3. Check where the data may go. If text cannot leave your systems, use an open-weight model on your own servers.
4. Run your own test. Write 50 to 100 real questions with the passage that should answer each, and count how often each model puts the right passage in its top few results.
5. Pick the smallest size that passes. Then store everything with that one model and that one size.
6. Plan for re-embedding. Models improve. Budget the time and cost to re-embed when you switch.

Common Failure Modes

A Short History of Embedding Models

How Embedding Models Show Up in Taskade

Embedding models run every time a Taskade agent searches its files. Files you attach to a Taskade AI agent as knowledge are searched by embedding similarity, so the agent finds the passages closest in meaning to a question. Taskade chooses and runs that embedding model, so you do not pick one or manage an index. Knowledge projects are read directly.

A policy lookup desk is the plain use. Upload your handbook, SOPs, and policy PDFs as files on an agent. A teammate asks a question in their own words, and the agent finds the closest passages and answers with a pointer to the source file. An automation logs each question and answer into a project, so you can see which questions the files do not answer yet. Taskade Genesis builds the first version of that app from one prompt, and Taskade EVE edits it in chat. To start from a working example, clone the Team Wiki Portal kit, an app kit with projects, an AI agent, and automations you can clone. The Free plan covers 3 apps and 1 agent and starts with 0 AI credits.

Describe yours and build it free →

Related Concepts

• Vector Embeddings: what an embedding is and why similar meanings land close together
• Latent Space: the learned map that embeddings are points in
• Vector Database: where embeddings are stored and searched
• Semantic Search: search by meaning, built on embeddings
• Retrieval-Augmented Generation: feeding retrieved passages to a language model
• Agentic RAG: retrieval that an agent plans and repeats
• Chunking: splitting documents into pieces worth embedding
• Representation Learning: how embedding models learn their features
• Platonic Representation Hypothesis: why different models still cannot be swapped freely
• Tokenizer: how text is split into the tokens that embedding prices count
• Agent Memory: how agents keep context over time
• Vector Databases Explained: the longer guide to vector search
• Latent Reasoning: models that reason in hidden vectors instead of words
• Sparse Autoencoders: a tool that names the features inside a model
• Steering Vectors: directions in latent space that change behavior

Frequently Asked Questions About Embedding Models

What is an embedding model?

An embedding model is a neural network that turns content such as a sentence, a document, or an image into a list of numbers called a vector embedding. Similar meanings get similar vectors, so software can find related content by measuring distance. Embedding models power semantic search, retrieval for AI agents, recommendations, and clustering.

How do I choose an embedding model?

Shortlist two or three models that fit your languages, content types, and input length. Then test them on 50 to 100 of your own questions and count how often the right passage appears near the top. Pick the smallest vector size that passes. Benchmarks such as MTEB help you build the shortlist, but your own test decides.

How many dimensions should an embedding have?

There is no single right size. Many 2026 models default to 1,024 to 3,072 dimensions, and many let you shrink them. Google's own figures for gemini-embedding-001 show its MTEB score drops from 68.17 at 1,536 dimensions to 67.99 at 768, while storage halves. Test the smallest size that still finds the right passages in your data.

How much does it cost to embed documents?

Usually little. On 2026-10-19, hosted text embedding prices ranged from $0.02 per million tokens for OpenAI's text-embedding-3-small and Voyage's voyage-4-lite to $0.20 for Google's gemini-embedding-2 at the standard rate. Embedding 10 million tokens once costs between $0.20 and $2.00 at those prices. Storage and search usually cost more over time.

What is MTEB?

MTEB, the Massive Text Embedding Benchmark, is the standard test suite for embedding models. The 2022 version covered 8 task types, 58 datasets, and 112 languages, and the 2025 MMTEB expansion grew it to over 500 tasks in 250+ languages. Its main lesson is that no single model wins every task, so treat a top score as a shortlist entry.

Can I mix embedding models in one index?

No. Each model places content at its own coordinates, so a query embedded with one model cannot meaningfully search vectors from another. Even two sizes of the same model can differ. Use one model and one size for both your documents and your queries, and re-embed everything when you change models.

Are open-weight embedding models good enough?

Often, yes. On the 2025 MMTEB benchmark, the best publicly available model was multilingual-e5-large-instruct with 560 million parameters. Open-weight models such as BGE-M3 (MIT license) and Qwen3-Embedding (Apache 2.0) run on your own hardware, which keeps your text inside your systems. You trade token fees for hardware and maintenance.

What is the difference between embeddings and an embedding model?

The embedding model is the network that does the conversion. Embeddings are its output, the vectors you store and search. One model produces embeddings for all of your content. If you switch models, the old embeddings are no longer comparable with new queries, so you create a new set of embeddings with the new model.

Further Reading

• OpenAI, Embeddings guide and pricing, read 2026-10-19
• Google, Gemini embeddings and pricing, read 2026-10-19
• Voyage AI, Embeddings and pricing, read 2026-10-19
• Cohere, Embed models, read 2026-10-19
• Muennighoff et al., MTEB: Massive Text Embedding Benchmark, 2022
• Enevoldsen et al., MMTEB: Massive Multilingual Text Embedding Benchmark, 2025
• Kusupati et al., Matryoshka Representation Learning, 2022
• MTEB leaderboard on Hugging Face