Summary: If you are picking an embedding API for a RAG app or semantic search, the three most common hosted choices price text embeddings within a narrow band of a few cents per million tokens, so cost is rarely the deciding factor. What really differs is the free allowance (Voyage gives every account a large free token pool, Cohere gives a free but non-commercial trial key, and OpenAI lists no free embedding allowance), the rate limits you start with, and a few billing rules that are easy to miss.
I checked the official pricing, rate-limit and model pages for Voyage AI, Cohere and OpenAI on 2026-11-09. Every number below comes from those pages; the cost scenarios are my own arithmetic on top of the listed prices, and I say so where that applies.
Why the embedding API choice matters more than the price tag
Embeddings are cheap per call, but they are sticky. Once you have embedded a corpus with one model, your vector database is full of vectors that only make sense next to other vectors from the same model (or the same compatible model family). Switching providers later means re-embedding everything and re-indexing. So the useful questions are:
- How much can I prototype before I pay anything?
- What happens to my bill and rate limits when I go from a test index to a full re-index?
- Are there model families that let me trade quality for speed without re-embedding?
- Are there billing rules (batch discounts, trial restrictions, data-use defaults) that change the real cost?
If you have not chosen where the vectors will live yet, my earlier comparison of Pinecone vs Qdrant vs Weaviate covers the database side, and LlamaParse vs Unstructured vs Docling covers turning PDFs into clean chunks before you embed them.
Voyage AI: a 200 million token free pool per account
Voyage (the docs header now reads "Voyage AI by MongoDB") bills text embeddings per token. According to its pricing page, the first 200 million tokens are free for every account on the current voyage-4 family (voyage-4-large, voyage-4, voyage-4-lite) and voyage-context-4, while some older domain models (voyage-finance-2, voyage-law-2, voyage-code-2) get 50 million free tokens. After the free pool, the listed prices per million tokens are:
- voyage-4-lite: $0.02
- voyage-4: $0.06
- voyage-4-large: $0.12
The Text Embeddings docs list a 32,000-token context length for all three voyage-4 models, a default of 1024 dimensions (with 256, 512 and 2048 also available through output_dimension), and state that all embeddings created with the 4 series are compatible with each other. In practice that means you could index documents with voyage-4-large and embed queries with the cheaper voyage-4-lite without rebuilding the index.
Voyage rate limits and the payment-method step
The rate-limits page lists "Basic" limits of 2,000 requests per minute for every model, with tokens-per-minute ranging from 3M (voyage-4-large) to 8M (voyage-4) and 16M (voyage-4-lite). Those numbers are the Tier 1 limits, and Tier 1 requires adding a payment method. Voyage notes that the free tokens still apply after you add a card. Tier 2 (at least $100 paid) doubles the limits and Tier 3 (at least $1,000 paid) triples them, and free-token usage does not count toward tier qualification.
Two Voyage billing rules worth knowing
- Batch API: Voyage's Batch API has a 12-hour completion window and a 33% discount, but the pricing page says free token credits do not apply to batch usage. If you are still inside your free 200M tokens, the regular endpoint is cheaper than the discounted batch endpoint.
- Data use: the FAQ says customers can opt out of Voyage storing and using their data for future model training, but opting out requires a payment method on file and an organization Admin, and you cannot opt back in from the dashboard afterwards.
Cohere: Embed 5 Pro and Fast, with a free trial key
On Cohere's pricing page (under the "Advanced retrieval models" tab) the current embedding models are listed as:
- Embed 5 Fast: $0.08 per 1M tokens
- Embed 5 Pro: $0.12 per 1M tokens
The model docs give the API names as embed-v5.0-pro and embed-v5.0-fast, a 128k-token context length, support for text, images and mixed text-and-image inputs such as PDF pages, and Matryoshka output dimensions from 256 up to 2048 (the default), as float, int8 or binary. Like Voyage's 4 series, Pro and Fast share an embedding space; Cohere explicitly recommends indexing with Pro and querying with Fast.
How the Cohere trial key really works
Cohere creates a Trial API key automatically when you sign up. Calls on a trial key are free, but the pricing FAQ says trial keys are rate limited and "not permitted to be used for production or commercial purposes". The rate-limits page adds that trial keys are limited to 1,000 API calls a month. For the Embed endpoint, both trial and production keys are listed at 2,000 inputs per minute; image embedding drops to 5 inputs per minute on a trial key versus 400 on production. To pay as you go you need a Production key, which an organization Owner requests through the "Go to Production" workflow on the Billing and Usage page. Production usage is billed at the end of each calendar month or whenever the outstanding balance reaches $250.
OpenAI: lowest listed price, no free embedding allowance
OpenAI's model pages list text-embedding-3-small at $0.02 and text-embedding-3-large at $0.13 per 1M tokens. The embeddings guide gives default vector lengths of 1536 (small) and 3072 (large), a dimensions parameter to shorten vectors, and a max input of 8,192 tokens for both models. That input limit is much shorter than Voyage's 32,000 or Cohere's 128k tokens, so long documents must be chunked before embedding, which you would usually do anyway for retrieval.
The model page shows rate limits by usage tier. For text-embedding-3-small, the Free tier row reads 100 requests per minute, 2,000 requests per day and 40,000 tokens per minute, while the Build tier jumps to 5,000 RPM and 1,000,000 TPM. I did not find a free token allowance for embeddings on these pages, so treat OpenAI embeddings as paid from the first token. One more limitation: the guide notes the third-generation models lack knowledge of events after September 2021, which matters only in edge cases for retrieval.
Side-by-side: free limits, prices and key features
| Voyage AI | Cohere | OpenAI | |
|---|---|---|---|
| Free allowance | First 200M tokens free per account (voyage-4 family); not usable for Batch API | Free trial key: 1,000 API calls/month, not for production or commercial use | No free embedding allowance listed |
| Price per 1M tokens | voyage-4-lite $0.02 · voyage-4 $0.06 · voyage-4-large $0.12 | Embed 5 Fast $0.08 · Embed 5 Pro $0.12 | text-embedding-3-small $0.02 · text-embedding-3-large $0.13 |
| Max input context | 32,000 tokens | 128k tokens | 8,192 tokens |
| Default dimensions | 1024 (256/512/2048 optional) | 2048 (256 to 1536 optional) | 1536 small / 3072 large (shorten with dimensions) |
| Starting rate limit | Tier 1 (card added): 2,000 RPM; 3M to 16M TPM by model | Embed: 2,000 inputs/min (trial and production) | Free tier: 100 RPM, 2,000 RPD, 40K TPM (3-small) |
| Mix models without re-indexing | Yes, voyage-4 series embeddings are compatible | Yes, Embed 5 Pro and Fast share a space | Not stated |
| Images | Separate multimodal models | Text, images and mixed inputs in Embed 5 | Text only |
Prices and limits from the official pages listed under Sources, checked 2026-11-09.
What a real indexing job costs (my arithmetic)
To make the per-token prices concrete, take a hypothetical knowledge base of 10,000 chunks averaging 800 tokens each, which is 8 million tokens per full index. Multiplying by the listed prices (excluding any free allowance):
- OpenAI text-embedding-3-small: 8 × $0.02 = $0.16
- Voyage voyage-4-lite: 8 × $0.02 = $0.16
- Voyage voyage-4: 8 × $0.06 = $0.48
- Cohere Embed 5 Fast: 8 × $0.08 = $0.64
- Cohere Embed 5 Pro or Voyage voyage-4-large: 8 × $0.12 = $0.96
- OpenAI text-embedding-3-large: 8 × $0.13 = $1.04
On Voyage, that same 8M-token index fits inside the 200M free pool 25 times over before you pay anything. The takeaway: for small and mid-size corpora, the embedding bill is usually smaller than your vector database or LLM bill. Quality on your own data, context length and rate limits should drive the choice, not a few cents per million tokens.
How to pick: a short decision path
- Prototyping with no budget: start with Voyage. The 200M free tokens cover many full re-indexes, and you only need a card to unlock Tier 1 limits.
- Free experiments, but not commercial: Cohere's trial key is free, but plan for the 1,000-calls-a-month cap and the non-commercial rule before you build anything customers will touch.
- Already paying OpenAI, short chunks, lowest price: text-embedding-3-small at $0.02 per 1M tokens is hard to beat on cost if your chunks fit in 8,192 tokens.
- Long documents or PDF pages with images: Cohere Embed 5 accepts 128k tokens and mixed text-and-image input in one model.
- Want to tune cost later without re-indexing: Voyage 4 series or Cohere Embed 5 Pro/Fast let you index with the larger model and query with the smaller one.
Whichever you choose, run the same small evaluation on all candidates: 50 to 100 real queries with known correct chunks, then compare recall at 5 or 10. If you want a repeatable way to do that kind of check, my Promptfoo tutorial shows how to set up automated test cases.
Editor's note
Who this is for: developers building a first RAG app, internal search or recommendation feature who need a hosted embedding API and want to understand free limits before signing up.
Not for: teams that must keep all data on-premises. Look at open-weight models you can self-host instead (Voyage's docs list an open-weight voyage-4-nano on Hugging Face).
Gotchas I found in the docs:
- Voyage free tokens do not apply to the Batch API, and tier upgrades count only billed usage, not free tokens.
- Opting out of Voyage training data use requires a payment method and cannot be undone from the dashboard.
- A Cohere trial key is free but not allowed for production or commercial use.
- OpenAI embedding inputs are capped at 8,192 tokens; longer text needs chunking first.
Common errors and fixes:
- HTTP 429 on Voyage: the docs recommend larger batches (a single request can hold up to 1,000 texts, within per-model token caps such as 320K for voyage-4 and 120K for voyage-4-large), a short sleep between requests, or exponential backoff.
- Poor search results after switching models: query vectors and document vectors must come from the same model or a compatible family; re-embed the corpus if you changed families.
- Mismatched vector length errors in your database: if you set
output_dimension(Voyage, Cohere) ordimensions(OpenAI), the index must be created with the same dimension. - Weak retrieval on Voyage or Cohere: set
input_typecorrectly:document/queryon Voyage,search_document/search_queryon Cohere.
FAQ
Is any of these embedding APIs completely free?
Not forever. Voyage gives each account the first 200 million tokens free on its current models, then charges per token. Cohere's trial key is free but capped at 1,000 calls a month and not allowed for commercial use. OpenAI lists no free embedding allowance.
Can I switch from OpenAI embeddings to Voyage or Cohere later?
Yes, but you have to re-embed every document with the new model and rebuild the index, because vectors from different model families are not comparable. Using the cost math above, re-embedding a few million tokens costs well under a dollar; the real cost is the engineering time and the re-indexing run.
Which is cheapest for a large one-off backfill?
By listed price, OpenAI text-embedding-3-small and Voyage voyage-4-lite are both $0.02 per 1M tokens. Voyage's Batch API takes 33% off its regular prices, though free tokens cannot be used there. Check quality on your own data before choosing purely on price.
Pick the provider whose free tier fits your prototype, test it on real queries, and switch only before your index gets big.
Last verified: 2026-11-09
Sources: Voyage AI Pricing, Voyage AI Rate Limits, Voyage AI Text Embeddings, Voyage AI FAQ, Cohere Pricing, Cohere API Keys and Rate Limits, Cohere Embed models, Cohere Introduction to Embeddings, OpenAI text-embedding-3-small, OpenAI text-embedding-3-large, OpenAI Embeddings guide.
Image credit: screenshots of the public Voyage AI pricing and rate-limit docs, the Cohere pricing page, and the OpenAI text-embedding-3-small model page, captured 2026-11-09. See the official screenshots at the end of this post.
댓글
댓글 쓰기