As of November 10, 2026, embedding 1 billion tokens costs $20 on OpenAI text-embedding-3-small, $130 on text-embedding-3-large, $100 on Mistral Embed and $200 on Gemini Embedding 2, on our arithmetic at list prices. Voyage AI's 200 million free tokens cut voyage-4-lite to $16, and Fireworks AI's smallest open models cost $8.
Text, code and multimodal embedding rates, free token allowances and batch discounts on nine providers, priced for a 1 billion token corpus and a 50 million token monthly refresh.
Creating the vectors is the cheapest line of a retrieval stack, and the price gap between models is still wider than most teams expect. OpenAI's own small and large embedding models differ by 6.5 times per token. Gemini Embedding 2 costs ten times OpenAI's small model for text, and more again for images, audio and video.
The real cost is rarely the first bill. It is the second one, when you switch models and have to re-embed everything, because vectors from one model family cannot be searched against vectors from another. This guide prices a full corpus, the monthly refresh, and that switch.
Toolradar data: our vector databases category covers 21 tools, where embeddings end up once they are created. 81% of those tools offer a free or freemium plan and 19% are paid only, so the vector store, not the embedding API, is usually where a retrieval budget is decided.
Methodology: nine embedding providers are compared on the per-token prices each vendor publishes, as of November 10, 2026, in US dollars. Totals are our arithmetic for text input at standard and batch rates, before taxes, volume contracts and the cost of storing the vectors. Benchmark quality scores are left out: retrieval quality depends on your own data, so test on a sample before committing.
Top Picks
Selected by pricing the same text corpus (1 million documents of about 1,000 tokens each) on each provider's published per-token embedding rate, then ranking on the one-off bill, the monthly refresh bill and how complete the public price list is. How we rate
Two text embedding models, a cheap small one and a stronger large one, on the API most teams already use.
Starting price
From $0.02 per 1M tokens
3.7 from 8 reviews
Best for
Teams already on OpenAI that want a cheap, well-documented default for text search.
Serverless inference platform that hosts open-weight embedding models, priced by model size.
Starting price
From $0.008 per 1M tokens
3.8 (17 reviews)
Best for
Cost-first teams happy to pick an open embedding model and run it on shared serverless capacity.
French AI lab with one general text embedding model and one built for code.
Starting price
$0.10 per 1M tokens (mistral-embed)
3.2 (150 reviews)
Best for
Teams that want a European vendor or a dedicated code embedding model.
| Tool | Starting price | Rating | Best for |
|---|---|---|---|
| OpenAI API Visit OpenAI API website | From $0.02 per 1M tokens | 3.7 from 8 reviews | Best for: Teams already on OpenAI that want a cheap, well-documented default for text search. |
| Fireworks AI Visit Fireworks AI website | From $0.008 per 1M tokens | 3.8 17 reviews | Best for: Cost-first teams happy to pick an open embedding model and run it on shared serverless capacity. |
| Mistral AI Visit Mistral AI website | $0.10 per 1M tokens (mistral-embed) | 3.2 150 reviews | Best for: Teams that want a European vendor or a dedicated code embedding model. |
| Google Gemini Visit Google Gemini website | $0.20 per 1M text tokens; free tier | 3.4 1,177 reviews | Best for: Products that search across images, audio or video as well as text, in one vector space. |
| Jina AI Visit Jina AI website | Free tokens on new API keys | 3.6 from 6 reviews | Best for: Teams that want an open-weight model they can later self-host without changing vectors. |
| Cohere Visit Cohere website | Check the current price on Cohere's site | 4.3 19 reviews | Best for: Enterprises that want embeddings, reranking and chat from one vendor with private deployment options. |
| Hugging Face Visit Hugging Face website | Inference Endpoints from $0.033 per hour | 3.6 11 reviews | Best for: Teams embedding steadily large volumes that prefer an hourly instance to per-token billing. |
| Amazon Bedrock Visit Amazon Bedrock website | Check the current price on AWS's site | 4.3 78 reviews | Best for: Teams whose data and vector store already live in AWS and must stay there. |
OpenAI API
Two text embedding models, a cheap small one and a stronger large one, on the API most teams already use.
- 3.7 on G2 (8 reviews)
- Starting price
- From $0.02 per 1M tokens
- Free plan
- No
- Rating
- 3.7 from 8 reviews
Best for: Teams already on OpenAI that want a cheap, well-documented default for text search.
- Pro: At list price it ties voyage-4-lite as the cheapest proprietary text model here.
- Pro: A dimensions parameter shortens 3,072-dimension vectors to save vector storage.
- Pro: OpenAI's guide says a 3-large vector cut to 256 dimensions still beats a full ada-002 vector.
- Con: Text only: no image, audio or code-specific embedding model.
- Con: The large model costs 6.5 times the small one, and inputs stop at 8,192 tokens.
The Batch API at 50% off and cached input discounts (up to 90% off) significantly lower costs for async or repetitive workloads.
Watch out
Overage charges at same per-token rates
Fireworks AI
Serverless inference platform that hosts open-weight embedding models, priced by model size.
- 4.1 on G2 (15 reviews)
- 2.9 on Trustpilot (2 reviews)
- Starting price
- From $0.008 per 1M tokens
- Free plan
- No
- Rating
- 3.8 (17 reviews)
Best for: Cost-first teams happy to pick an open embedding model and run it on shared serverless capacity.
- Pro: The lowest text rate here: $8 for 1 billion tokens on models up to 150M parameters.
- Pro: Models of 150M to 350M parameters cost $0.016, so the bill doubles with model size.
- Pro: The same account serves chat and other open models on one bill.
- Con: Price is set by parameter band, so you choose and evaluate the open model yourself.
- Con: No free token allowance is printed for embeddings.
There is no free plan.
Watch out
US-only serverless carries a 50% premium on select models
Mistral AI
French AI lab with one general text embedding model and one built for code.
- 2.4 on Trustpilot (95 reviews)
- 4.2 on G2 (55 reviews)
- Starting price
- $0.10 per 1M tokens (mistral-embed)
- Free plan
- Yes
- Rating
- 3.2 (150 reviews)
Best for: Teams that want a European vendor or a dedicated code embedding model.
- Pro: Codestral Embed is built for code retrieval, a gap in OpenAI's lineup.
- Pro: Embedding and batch endpoints sit in the same API as Mistral's chat models.
- Pro: Each model page shows its price in USD or EUR.
- Con: mistral-embed costs five times OpenAI's small model for general text.
- Con: Context stops at 8k tokens, so long documents need chunking.
Developers pay per token on the API, with 50% off batch jobs.
Google Gemini
Gemini API embedding models, including a multimodal one covering text, images, audio, video and PDFs.
- 4.4 on G2 (610 reviews)
- 1.5 on Trustpilot (503 reviews)
- 4.6 on Capterra (64 reviews)
- Starting price
- $0.20 per 1M text tokens; free tier
- Free plan
- Yes
- Rating
- 3.4 (1,177 reviews)
Best for: Products that search across images, audio or video as well as text, in one vector space.
- Pro: Batch mode halves the price of both embedding models.
- Pro: Output dimensions are flexible from 128 to 3,072, so storage can be tuned.
- Pro: The free tier charges nothing for embedding text, images, audio or video.
- Con: Free-tier data is used to improve Google's products; the paid tier opts out.
- Con: Embedding 2 text costs ten times OpenAI's small model, and video costs $12.00 per million tokens.
Developer access starts free in Google AI Studio and moves to pay by usage.
Jina AI
Embedding and reranker specialist with open-weight v5 models sold through a prepaid token API.
- 3.0 on SourceForge (2 reviews)
- 4.5 on G2 (2 reviews)
- 3.2 on Trustpilot (2 reviews)
- Starting price
- Free tokens on new API keys
- Free plan
- Yes
- Rating
- 3.6 from 6 reviews
Best for: Teams that want an open-weight model they can later self-host without changing vectors.
- Pro: jina-embeddings-v5-text comes in 677M small and 239M nano sizes.
- Pro: The small model reads 32K tokens of context, the nano model 8K.
- Pro: Every new API key gets free tokens usable across Jina's models.
- Con: Tokens are sold as prepaid top-ups, so compare the per-token cost at checkout.
- Con: Buying credit up front suits steady use less well than a monthly invoice.
- Starting price
- Check the current price on Cohere's site
- Free plan
- Yes
- Rating
- 4.3 (19 reviews)
Best for: Enterprises that want embeddings, reranking and chat from one vendor with private deployment options.
- Pro: Embed 5 is its newest family of embedding models.
- Pro: Rerank, its retrieval optimization model, is sold alongside Embed.
- Pro: A trial key gives access to all models for proofs of concept.
- Con: Its pricing page lists Command and Model Vault rates but no Embed price.
- Con: Without a public rate, a corpus bill cannot be compared before a sales call.
The free trial key is generous for testing, but Model Vault's GPU-hour costs can add up for heavy use.
Hugging Face
Model hub that rents dedicated, autoscaling endpoints to run any open embedding model.
- 2.6 on Trustpilot (7 reviews)
- 4.9 on G2 (4 reviews)
- Starting price
- Inference Endpoints from $0.033 per hour
- Free plan
- Yes
- Rating
- 3.6 (11 reviews)
Best for: Teams embedding steadily large volumes that prefer an hourly instance to per-token billing.
- Pro: A 1 vCPU, 2 GB Intel Sapphire Rapids instance lists at $0.03 an hour, $24 a month.
- Pro: Any open embedding model on the Hub can be deployed, such as Voyage AI's open-weight voyage-4-nano.
- Pro: Fixed hourly cost makes very high token volumes predictable.
- Con: You size, monitor and scale the instance yourself; idle hours still bill.
- Con: For a one-off corpus, per-token APIs are usually simpler and cheaper.
Best for developers, researchers, and teams wanting a central AI hub without high upfront costs.
Watch out
Inference endpoints from $0.033 per hour
Amazon Bedrock
AWS's managed model service, with Amazon Titan Text Embeddings V2 and Nova Multimodal Embeddings.
- 4.3 on G2 (78 reviews)
- Starting price
- Check the current price on AWS's site
- Rating
- 4.3 (78 reviews)
Best for: Teams whose data and vector store already live in AWS and must stay there.
- Pro: Titan Text Embeddings V2 and Nova Multimodal Embeddings run inside your AWS account.
- Pro: Batch pricing is offered for Nova Multimodal Embeddings.
- Pro: Billing joins your existing AWS invoice and IAM controls.
- Con: Prices are set per AWS region, so confirm the rate for yours.
- Con: Fewer embedding model choices than a specialist such as Voyage AI.
Amazon Bedrock's On-Demand pricing is fair for initial exploration, offering pay-per-use flexibility without commitment.
Watch out
Storage costs for fine-tuning data
Other Vector Databases tools worth considering
More published tools from our Vector Databases category, ordered by our category ranking. They are not part of the editorial picks above.
What an embedding API charges for
An embedding API turns a piece of text (or an image, a clip of audio, a code file) into a list of numbers that captures its meaning, so a vector database can find similar items. Retrieval-augmented generation, semantic search, deduplication and recommendations all start here.
Almost every provider bills one meter: input tokens, priced per million. There is no output charge, because the output is the vector itself. Three details change the bill:
| Bill detail | What it means | Example |
|---|---|---|
| Model tier | Small, standard and large models from the same vendor | 6.5 times apart on OpenAI's small and large models |
| Batch rate | Asynchronous jobs at a discount, with a completion window | Voyage AI takes 33% off with a 12-hour window |
| Free allowance | Tokens or a tier that costs nothing | Voyage AI's first 200 million tokens per account |
| Input type | Multimodal models price images, audio and video separately | $0.45 per million image tokens on Gemini Embedding 2 |
OpenAI's embeddings guide works on about 800 tokens per page of text, so 1 billion tokens is about 1.25 million pages on our arithmetic.
How much does it cost to embed 1 billion tokens?
The scenario: 1 million documents of about 1,000 tokens each, so 1 billion tokens embedded once, then 50 million tokens of new or changed content embedded every month. Rates as of November 10, 2026; totals are our arithmetic at list prices for text.
| Model | Per 1M tokens | Embed 1B tokens | 50M tokens a month | Batch, 1B tokens |
|---|---|---|---|---|
| Fireworks AI, models up to 150M parameters | $0.008 | $8 | $0.40 | Check Fireworks AI's site |
| Voyage AI voyage-4-lite | $0.02 | $16 after free tokens | $1 | $13.40 |
| OpenAI text-embedding-3-small | $0.02 | $20 | $1 | Check OpenAI's site |
| Voyage AI voyage-4 | $0.06 | $48 after free tokens | $3 | $40.20 |
| Voyage AI voyage-4-large | $0.12 | $96 after free tokens | $6 | $80.40 |
| Mistral Embed | $0.10 | $100 | $5 | Check Mistral's site |
| OpenAI text-embedding-3-large | $0.13 | $130 | $6.50 | Check OpenAI's site |
| Gemini Embedding 001 | $0.15 | $150 | $7.50 | $75 |
| Gemini Embedding 2 | $0.20 | $200 | $10 | $100 |
Voyage AI's free allowance is a one-time credit, not a monthly one. The first 200 million tokens per account are free on its 4-series models, so the 1 billion token corpus pays for 800 million. The monthly refresh then pays full price. Free tokens do not apply to its Batch API, which still comes out cheaper at this volume: 33% off every token beats 20% of tokens free.
At this scale the embedding bill is small next to the people cost of the project. Even the dearest text model here costs $200 for the whole corpus. The decision that matters is which model you can live with for years, since leaving it means paying the corpus bill again.
Key Features to Look For
Price per million input tokens (Essential)
Text rates span 25 times from Fireworks AI's smallest open models to Gemini Embedding 2, so the model choice moves the bill more than the vendor does.
Free allowance (Essential)
Voyage AI gives each account 200 million free tokens on its 4-series models; Gemini's free tier costs nothing but lets Google use the data to improve its products.
Batch discount (Important)
Gemini halves its embedding price for batch jobs and Voyage AI takes 33% off with a 12-hour completion window, but Voyage's free tokens do not apply to batch usage.
Output dimensions (Important)
OpenAI returns 1,536 or 3,072 dimensions by default, Voyage AI 1,024, and Gemini Embedding 2 anything from 128 to 3,072; fewer dimensions means a smaller vector database bill.
Context length (Important)
OpenAI and Gemini Embedding 2 accept 8,192 input tokens, Mistral Embed 8k, and Voyage AI's 4-series models 32,000, which decides how you chunk long documents.
Model compatibility (Important)
Voyage AI says all 4-series embeddings are compatible with each other, so you can move between lite and large without re-embedding; across vendors, vectors never transfer.
Multimodal input (Nice to have)
Gemini Embedding 2 maps text, images, audio, video and PDFs into one space, at $6.50 per million audio tokens and $12.00 per million video tokens.
What to settle before you pick an embedding model
-
Count your corpus in tokens, not documents: run a sample through the vendor's tokenizer and multiply.
-
Estimate monthly churn, the share of documents added or edited, since that is your recurring bill.
-
Decide on dimensions with your vector database bill in mind; a 3,072-dimension vector takes three times the space of a 1,024 one.
-
Check whether your data can be used for training: Gemini's free tier says yes, its paid tier says no.
-
Pick a model family you expect to keep, because switching means re-embedding the full corpus.
-
If you index code, images or audio, price the specialized or multimodal model, not the text rate.
Evaluation Checklist
Embed a few thousand of your own documents on two or three candidate models and compare search results on real queries.
Measure your average tokens per chunk with each vendor's tokenizer; tokenizers differ, so the same text can bill differently.
Price the one-off corpus at the batch rate where one exists, and the daily trickle at the standard rate.
Work out vector storage at your chosen dimensions and precision before you commit to a large model.
Read the data-use terms of any free tier before sending customer data through it.
Confirm rate limits for a bulk backfill, which can take far longer than the bill suggests.
Pricing Overview
Free allowance
Prototypes and small corpora: Voyage AI's 200 million free tokens per account, Gemini's free tier, and the free tokens Jina AI puts on every new API key.
$0
Small and lite models
High-volume text search where cost matters more than the last point of retrieval quality: OpenAI text-embedding-3-small, voyage-4-lite and Fireworks AI's small open models.
From $0.008 per 1M tokens
Standard and large models
Multilingual or harder retrieval: voyage-4 and voyage-4-large, Mistral Embed, OpenAI text-embedding-3-large and both Gemini embedding models.
$0.06 to $0.20
Batch jobs
One-off backfills that can wait hours: Gemini batch at half price, Voyage AI batch with a 12-hour window.
33% to 50% off standard
Dedicated endpoints
Running an open-weight embedding model on your own instance through Hugging Face Inference Endpoints, billed by the hour instead of per token.
From $0.033 per hour
Enterprise and quotes
Private deployments, fine-tuned models and committed volume, as offered by Cohere, Voyage AI and Jina AI.
Contact sales
Pricing Comparison
| Provider | Text model and rate per 1M tokens | Batch rate | Free allowance | Embed 1B tokens |
|---|---|---|---|---|
OpenAI | text-embedding-3-small $0.02; text-embedding-3-large $0.13 | Check the current price on OpenAI's site | None | $20 small; $130 large |
Voyage AI | voyage-4-lite $0.02; voyage-4 $0.06; voyage-4-large $0.12 | 33% off, 12-hour window | 200M tokens per account | $16, $48 or $96 |
gemini-embedding-001 $0.15; gemini-embedding-2 $0.20 | $0.075; $0.10 | Free tier, data used to improve products | $150 or $200 | |
mistral-embed $0.10; codestral-embed $0.15 | Check the current price on Mistral's site | None | $100 text; $150 code | |
Up to 150M parameters $0.008; 150M to 350M $0.016; Qwen3 8B $0.10 | Check the current price on Fireworks AI's site | None | $8, $16 or $100 | |
Inference Endpoints, billed per instance hour | Not applicable | None | Depends on instance hours | |
Cohere | Check the current price on Cohere's site | Check the current price on Cohere's site | Trial API key | Not calculated |
Check the current price on Jina AI's site | Check the current price on Jina AI's site | Free tokens on new API keys | Not calculated | |
Check the current price on AWS's site | Check the current price on AWS's site | None | Not calculated |
Prices as of November 10, 2026 from OpenAI text-embedding-3-small, OpenAI text-embedding-3-large, Voyage AI pricing, Gemini API pricing, Mistral Embed, Codestral Embed, Fireworks AI pricing, Hugging Face pricing, Cohere pricing, Jina AI embeddings and Amazon Bedrock pricing. Totals are our arithmetic for 1 billion text tokens at standard rates; Voyage AI totals subtract its 200 million free tokens. Dimensions and context limits come from the OpenAI embeddings guide, Voyage AI embeddings docs and the Gemini Embedding 2 model page.
Mistakes to Avoid
- ×
Choosing the largest model by default: OpenAI's large model costs 6.5 times its small one for gains you may not see on your data.
- ×
Treating Voyage AI's 200 million free tokens as monthly; they are a one-time allowance per account.
- ×
Running a big backfill at the standard rate when a batch rate halves the bill on Gemini.
- ×
Storing full 3,072-dimension float vectors when 1,024 dimensions would serve the same queries.
- ×
Switching embedding vendors without budgeting for re-embedding the whole corpus.
Expert Tips
- →
Use batch rates for the backfill and standard rates for live updates; both Gemini and Voyage AI price the two differently.
- →
Shrink vectors before they hit storage. At float32, 1 million vectors take about 12.3 GB at 3,072 dimensions and 4.1 GB at 1,024, on our arithmetic; Voyage AI's int8 output at 1,024 dimensions cuts that to about 1 GB.
- →
Stay within one model family where you can. Voyage AI's 4-series vectors are compatible across sizes, and Jina says its v5-omni models are byte-for-byte compatible with v5-text, with no reindexing.
- →
Compare the chat and embedding sides of a vendor together: see Mistral AI vs OpenAI API or Cohere vs Mistral AI, and OpenAI API pricing for the rest of the rate card.
- →
Choosing where the vectors live next? Start with the best vector databases, the vector database statistics and the NLP tools category.
Red Flags to Watch For
-
A free tier whose terms let the vendor use your data, used on customer documents.
-
An embedding price that is not printed anywhere public, when the corpus bill is your main cost.
-
A model announced without a stated deprecation policy, since a retired model forces a full re-embed.
-
Benchmark claims with no test on your own documents and queries.
The Bottom Line
For most text search, OpenAI's text-embedding-3-small is the default worth beating: a full 1 billion token corpus costs $20, and its dimensions parameter keeps storage in check. Voyage AI is the strongest alternative on price: its free allowance and 33% batch discount undercut OpenAI's small model with voyage-4-lite, and its larger models stay compatible with the lite one.
Fireworks AI is the cheapest per token if you are willing to choose an open model, and Google Gemini is the pick when images, audio or video need to sit in the same index as text. Whatever you choose, budget for re-embedding once: it is the cost that surprises teams, not the first bill. See the OpenAI API alternatives and best RAG tools for the rest of the stack.
Frequently Asked Questions
How much does it cost to embed 1 million documents in 2026?
For 1 million documents of about 1,000 tokens each, 1 billion tokens in all, our arithmetic gives $20 on OpenAI text-embedding-3-small, $130 on text-embedding-3-large and $200 on Gemini Embedding 2, at list prices as of November 10, 2026. Voyage AI's voyage-4-lite comes to $16 after its 200 million free tokens.
What is the cheapest embedding API?
On published per-token rates, Fireworks AI is cheapest at $0.008 per million tokens for open models up to 150M parameters. Among proprietary models, OpenAI text-embedding-3-small and Voyage AI voyage-4-lite share the same per-token rate, and Voyage's batch rate takes a further 33% off.
Is there a free embedding API?
Yes, with limits. Gemini's free tier charges nothing for embeddings, but Google may use that data to improve its products. Voyage AI gives each account 200 million free tokens on its 4-series models, and Jina AI puts free tokens on every new API key.
Is OpenAI text-embedding-3-large worth the price over 3-small?
Only if it measurably improves retrieval on your own data. It costs $0.13 per million tokens, 6.5 times the small model's rate, and returns 3,072 dimensions instead of 1,536, which also doubles vector storage unless you shorten it with the dimensions parameter.
How much does Gemini embedding cost?
Gemini Embedding 2 costs $0.20 per million text tokens on the paid tier, $0.45 for images, $6.50 for audio and $12.00 for video, with batch mode at half price. The text-only gemini-embedding-001 costs $0.15, or $0.075 in batch. The free tier costs nothing but uses your data to improve Google's products.
Do I have to re-embed my data if I switch embedding models?
Yes. Vectors from different model families are not comparable, as Jina AI's own documentation puts it, so a switch means embedding the full corpus again on the new model. Within a family it can differ: Voyage AI says all its 4-series embeddings are compatible with each other.
Does Voyage AI's free allowance renew each month?
No. Voyage AI's pricing page describes the first 200 million tokens on its 4-series models as free for every account, a one-time allowance. Batch API usage does not draw on it and is billed at the discounted batch rate instead.
Cite this page: Toolradar, "Embedding API Pricing Compared (2026)", updated November 2026, https://toolradar.com/guides/embedding-api-pricing-compared
Spotted an outdated price or a wrong fact? Suggest a correction: we check it against the vendor's own page. Work for one of these vendors? Claim your listing to update it yourself.
Related Guides
Ready to Choose?
Compare features, read reviews, and find the right tool.
Some offers on this page may be paid placements or contain affiliate links.