Embedding Models Compared: 10,000 Documents Cost $8 or Less to Embed

Embedding a 10,000-document corpus costs between $0.80 and $8.00 at the published rate cards of OpenAI, Cohere, Voyage AI and Google. That spread is real, and it is also small. An embedding models comparison that starts with price starts in the wrong place, because the model sets a ceiling on what your retrieval can find. Nearly every score that measures that ceiling was published by the vendor selling it.

What an embedding model decides in a RAG system

An embedding model turns a passage of text into a vector, a fixed-length list of numbers. That lets a database find passages whose meaning sits close to a query. In retrieval-augmented generation, those retrieved passages are everything the language model gets to read before it answers.

That makes the embedding model a ceiling rather than a dial. A reranker can reorder what the first stage found. A better prompt can use the results more carefully. Neither one can recover a passage that the first stage never returned.

Most teams still pick one by default. They use the model from their framework's quickstart, or the one sold by the vendor already on their invoice. I do not think defaulting is the mistake. The mistake is defaulting without writing down what you chose, what it scored on your own data, and what it would cost to change.

The choice fixes more than retrieval quality. It sets the maximum length of each chunk you can embed, the number of dimensions every vector carries, and a vector space that no other model can read. That last property is why the decision is harder to reverse than its small price suggests. The separate argument for pairing vectors with keyword search is laid out in the review of hybrid search and reranker benchmarks, and it applies whichever model you pick here.

Embedding models comparison: the rate card table

The table below lists the hosted text embedding models from the four vendors most RAG buyers shortlist. Every price is the vendor's own published rate per 1 million input tokens, checked against the vendor page on 29 October 2026. Embedding models bill input only, because they produce no generated text.

Sources: OpenAI API pricing, Voyage AI pricing, Cohere's Embed 5 launch post and embed documentation, and Google's Gemini API pricing and embeddings pages. Cohere's public pricing page lists hourly dedicated-hosting rates for Embed 5 but no per-token rate for embed-v4.0, so that cell is left empty rather than filled from a third-party tracker.

Three spreads stand out. Price runs from $0.02 to $0.20, a 10 times gap. Maximum input runs from 2,048 tokens on gemini-embedding-001 to 128,000 on Cohere's models, a 62 times gap. Default dimensions run from 1,024 to 3,072, a 3 times gap that lands directly on your storage bill.

Of the three, price matters least for most corpora, and input length matters most for documents that cannot be split cleanly. A contract, a long support thread or a technical specification often loses meaning when it is cut into 2,048-token pieces. If your documents look like that, context length is a selection criterion before score is.

One honest gap in the table: OpenAI's pricing page does not list a separate batch rate for its embedding models, and its general note says the Batch API is billed at the chosen model's rates. Do not assume a 50% batch discount on OpenAI embeddings unless your own invoice shows one. Voyage states a 33% batch discount and Google lists batch at half the standard rate.

What it costs to embed a 10,000-document corpus

A per-token rate is hard to reason about until it is attached to a corpus. So here is one, priced across every model in the table that has a published rate.

The assumption behind the number

Assume 10,000 documents averaging 4,000 tokens each. That is 5 pages per document at the 800-tokens-per-page convention OpenAI uses in its own embeddings guide. The corpus totals 40 million tokens. A knowledge base of policy documents, product manuals and support articles at a mid-sized company sits comfortably in that range.

Two adjustments apply in practice. Chunk overlap, where each chunk repeats the tail of the previous one, inflates the token count; at 20% overlap, multiply every figure below by 1.2. And each vendor uses its own tokenizer, so the same 10,000 documents will not count as exactly 40 million tokens everywhere.

Calculation: 40 million tokens multiplied by each list price per 1 million tokens. Voyage's 200 million free tokens apply per account to each Voyage 4 model at standard rates, not to batch. Cohere image pricing is excluded because the corpus is assumed to be text.

Batch rates and free tokens reshuffle the order

List price is not what a careful buyer pays. Google's batch rate halves both Gemini models. Voyage gives each account 200 million free tokens on each Voyage 4 model, which covers this corpus five times over before the first invoice.

Query embedding barely moves the total. Every search embeds the query too, but queries are short. One million queries at an illustrative 30 tokens each is 30 million tokens, or $3.60 on a $0.12 model. The volume that changes the economics is re-embedding the corpus, which is covered below.

My position is simple. A $5 difference should not choose the model that sets your retrieval ceiling for the next two years. If your corpus is in the tens of millions of tokens, ignore price until the final tie-break.

Why every MTEB leaderboard number here is self-reported

MTEB, the Massive Text Embedding Benchmark, is the standard public yardstick for embedding models. Its community expansion, MMTEB, covers over 500 quality-controlled tasks across 250-plus languages. The live MTEB leaderboard is a web application whose rankings shift as models and task sets are added, and it does not offer a citable dated snapshot. So this section relies on the tables each vendor published, with dates, and says who measured what.

Google's paper: a 9-point multilingual lead

Google's March 2025 technical report on Gemini Embedding, the model released as gemini-embedding-001, reports a mean task score of 68.32 on multilingual MTEB. In the same table, OpenAI's text-embedding-3-large scored 58.92 and Cohere's older multilingual v3 model scored 61.10. On the retrieval subset, the scores were 67.71 for Gemini and 59.27 for OpenAI.

The evaluation covered 164 tasks: 132 multilingual, 41 English and 12 code retrieval. That is a serious piece of work. It is also Google scoring Google, on the task mix Google chose.

OpenAI's guide: 64.6, on an older yardstick

OpenAI's embeddings guide lists text-embedding-3-large at 64.6% on its MTEB evaluation, text-embedding-3-small at 62.3% and the older ada-002 at 61.0%. It compares the three models only with each other. Google's paper measured the same text-embedding-3-large at 66.43 on MTEB English v2.

Neither number is wrong. They come from different versions of the benchmark with different task sets. The lesson is that an MTEB score without a version and a task subset is not a comparable number, and most vendor slides omit both.

Voyage on RTEB: 14.05% over OpenAI

Voyage's January 2026 launch post scores voyage-4-large across all 29 datasets in the Retrieval Embedding Benchmark (RTEB). It reports average margins of 3.87% over Gemini Embedding 001, 8.20% over Cohere Embed v4 and 14.05% over OpenAI v3 Large. The metric is normalised discounted cumulative gain on the top 10 results, which rewards putting the right passage near the top.

The post does not state whether those margins are relative or absolute. Its earlier voyage-3.5 post, from May 2025, claimed an 8.26% edge over OpenAI v3 Large across 100 datasets in 8 domains. Both are Voyage's evaluations on Voyage's selection.

Cohere on ViDoRe V3: 85.8 against 75.5

Cohere launched Embed 5 on 30 September 2026. On the ViDoRe V3 document retrieval benchmark, it reports Embed 5 Pro at 85.8, voyage-4-large at 83.7, Gemini Embedding 2 at 83.2 and OpenAI text-embedding-3-large at 75.5. The scores use Cohere's own metric variant, which it calls RCP-nDCG@10.

ViDoRe tests retrieval over visually rich documents such as reports and slides. That suits a multimodal model priced at $0.40 per million image tokens. It says little about plain-text support tickets.

The pattern is not fraud. It is selection. Each vendor chooses the benchmark where its model was designed to be strongest, and each result can be true at the same time. The broader version of this problem, across language model leaderboards, is examined in the analysis of why LLM leaderboards disagree. For a buyer, the consequence is direct: no published number here predicts how any of these models will rank on your documents.

The costs that outlast the embedding bill

The embedding invoice is a one-off for a static corpus. Two other costs recur, and both are set by the model you pick on day one.

Dimensions set your storage line

Each vector is stored as one number per dimension, usually a 4-byte float. Split the 40 million-token corpus into 512-token chunks and you get 78,125 vectors. At 1,024 dimensions they occupy 320 MB. At 3,072 dimensions, the default for both Gemini models and text-embedding-3-large, they occupy 960 MB.

That triples the memory a vector database must hold for fast search, and the multiple carries through to every replica and every index rebuild. What that memory costs on each database is set out in the vector database cost comparison.

Most of these models now let you shorten vectors. Google's own documentation is unusually useful here: it lists gemini-embedding-001 at 68.16 on MTEB at 2,048 dimensions and 67.99 at 768 dimensions. On Google's numbers, cutting to 768 dimensions costs 0.17 points and saves 75% of the storage of the 3,072 default.

Re-embedding is the real lock-in

Vectors from one model are meaningless to another. Switching vendors, or even moving between generations from the same vendor, means re-embedding the full corpus, rebuilding the index and re-running your evaluation. The token bill for that is the small part. The engineering time and the risk of a silent quality drop are the large part.

Voyage is the one vendor in this set addressing this directly. MongoDB's product announcement for the Voyage 4 series says the four models share a compatible embedding space, so you can index with voyage-4-large and query with a cheaper sibling. That compatibility does not extend to voyage-3.5 or earlier models. The general shape of switching costs between model vendors is covered in the breakdown of model switching costs.

Where this comparison is weakest

There is no independent, dated score in this post that covers all four vendors on one task set. The live MTEB leaderboard comes closest, and it could not be captured as a fixed, citable snapshot. Every quality figure above is a vendor publication, and should be read as directional.

The cost model rests on one assumption: 4,000 tokens per document. Your corpus could be ten times larger or smaller. Tokenizers also differ, so 40 million OpenAI tokens is not exactly 40 million Voyage or Gemini tokens. The ranking in the cost table should survive that; the absolute numbers will not.

Open-weight models are excluded. The MMTEB authors found that the best publicly available model on their benchmark was multilingual-e5-large-instruct, at only 560 million parameters, and Voyage now publishes an open-weight voyage-4-nano. If you already run GPU capacity, self-hosting can beat every row in the price table.

Finally, the case for choosing on price is stronger than this post's headline admits, at scale. Embedding 10 billion tokens costs $200 at $0.02 per million and $2,000 at $0.20. A team indexing the open web, or re-embedding nightly, should weigh price far earlier than a team indexing a knowledge base.

Best embedding model for RAG: pick in this order

There is no best embedding model for RAG in general. There is a best one for your documents, languages and queries, and the only way to find it is to test on those. Here is the order I would apply the filters in.

Step 3 carries the weight. Fifty real queries, each paired with the passage a human says should come back, is an afternoon's work. Building that set properly is covered in the guide to building an LLM evaluation suite. If your documents are short enough to skip retrieval altogether, check the cost comparison of RAG against long context before you choose any embedding model.

If I were starting an English-language text RAG build this week, I would test voyage-4, text-embedding-3-small and Cohere Embed 5 Fast against each other. That covers three vendors at $0.02 to $0.08 per million tokens. I would only move up to a flagship if it won my own test by a margin I could see in the answers.

Frequently asked questions

Where to start this week

Pull 50 real questions from your support inbox, search logs or internal chat, and next to each one write the document that should answer it. That list is the only benchmark in this post that was written about your data, and it will outlive whichever model you pick.

Then open your vector store and write down the model name and dimension count behind your current index. If the answer is a 3,072-dimension default that nobody chose, test a 768 or 1,024-dimension setting against your 50 questions before your next re-index.

Related reading