Voyage AI

Editor’s note: This Voyage AI review is based on my hands-on test of the Voyage AI API, including voyage-4-large, voyage-4, voyage-4-lite, rerank-3 and the Batch API, in November 2026. I weighed pricing and the free-token allowance, the voyage-4 models and their shared embedding space, rerankers and contextualized chunk embeddings, the MongoDB Atlas route, rate limits and data controls, and how Voyage compares with OpenAI, Google Gemini, Cohere, Jina and open-weight models.

Quick verdict: I recommend Voyage AI for developers building RAG, semantic search or coding-agent retrieval who want 32,000-token inputs, cheap per-token pricing and a reranker from the same vendor. The trade-off: it is a bare model API (no hosted index, no self-serve fine-tuning), and it may train on your API data until an org Admin with a payment method opts out, a switch that cannot be reversed in the dashboard.

Voyage AI, owned by MongoDB since February 2025, sells embedding models and rerankers through one API. Its four voyage-4 models share one embedding space, so documents embedded with one can be searched with queries from another. voyage-4-lite costs $0.02 per 1M tokens, and every account gets 200M free tokens per current model, making entry pricing a strong point in my Voyage AI review.

In this review, I’ll take a closer look at Voyage AI’s pricing, features, and MongoDB integration, so you can see exactly whether it fits your retrieval stack or whether OpenAI or Cohere embeddings are the better buy.

Key Takeaways

• Shared embedding space: index with voyage-4-large, query with voyage-4-lite, no re-embedding.
• Low per-token prices: $0.02 (voyage-4-lite), $0.06 (voyage-4) and $0.12 (voyage-4-large) per 1M tokens, 200M free per model, 33% off on batch.
• Reranker on the same key: rerank-3 ($0.05 per 1M) and rerank-3-lite ($0.02) take 32,000-token inputs and up to 1,000 documents per call.
• Training opt-out is off by default: it needs an org Admin plus a payment method, and the dashboard cannot undo it.
• No uptime credits, no backward compatibility: 99.5% uptime is a service level objective with no credits, and voyage-3.x to voyage-4 means re-embedding the corpus.

Pros and Cons

Voyage AI Pricing: How Much Does It Cost?

Voyage AI has no subscription plans: every model is pay as you go per token, billed monthly, with the card charged around the 2nd of the month. Each account gets the first 200M tokens free on each current model (50M on voyage-law-2, voyage-finance-2 and voyage-code-2).

• voyage-4-lite ($0.02 per 1M tokens): for high-volume query embedding and latency-sensitive apps, with 16M tokens per minute at Tier 1.
• voyage-4 ($0.06 per 1M tokens): for most RAG and semantic search.
• voyage-4-large ($0.12 per 1M tokens): for corpus embedding where retrieval quality matters most.
• voyage-code-4 and voyage-context-4 ($0.12 per 1M tokens): for code retrieval and coding agents, and for chunk embeddings that carry whole-document context.
• rerank-3 ($0.05) and rerank-3-lite ($0.02 per 1M tokens): for reordering retrieved candidates; a 100-document rerank-3 request at 500 tokens each costs about $0.0025.
• voyage-multimodal-3.5 ($0.12 per 1M text tokens plus $0.60 per billion pixels): for PDF screenshots, slides and video; one image costs between $0.00003 and $0.0012.

Model	Price per 1M tokens	Free allowance	Best for
voyage-4-lite	$0.02	200M tokens	High-volume queries
voyage-4	$0.06	200M tokens	General RAG
voyage-4-large	$0.12	200M tokens	Corpus embedding
voyage-code-4	$0.12	200M tokens	Code retrieval
voyage-context-4	$0.12	200M tokens	Document-aware chunks
voyage-multimodal-3.5	$0.12 text + $0.60/billion pixels	200M tokens + 150B pixels	Visual retrieval
rerank-3	$0.05	200M tokens	Candidate reranking
rerank-3-lite	$0.02	200M tokens	Lower-cost reranking
voyage-law-2 / voyage-finance-2	$0.12	50M tokens each	Existing domain workloads

The Batch API takes 33% off with a 12-hour completion window and up to 100K inputs per batch, but free tokens do not apply to batch usage. Files API storage costs $0.05 per GB per month, and files are deleted after 30 days.

Older voyage-3-large, voyage-code-3 and voyage-context-3 cost $0.18 per 1M tokens with no free tokens, so staying on them costs 50% more than voyage-4-large. Fine-tuned models on dedicated instances are sales-only.

• Tier 1: adding a payment method gives 2,000 requests per minute on every model and 2M to 16M tokens per minute, depending on the model.
• Tier 2: $100 paid doubles Tier 1 limits; buying $100 of prepaid credits jumps straight here.
• Tier 3: $1,000 paid triples Tier 1 limits. Free tokens do not count toward either tier.

Atlas charges the same list prices, but the key decides the bill: an Atlas-created key bills through Atlas, a Voyage-created key through Voyage. Pinning Automated Embedding inference to the US or EU costs 1.1x.

Is Voyage AI Good Value for Money?

• Entry-price parity: voyage-4-lite matches OpenAI text-embedding-3-small at $0.02 per 1M tokens.
• Top tier undercuts rivals: voyage-4-large at $0.12 sits just under OpenAI text-embedding-3-large at $0.13 and below gemini-embedding-001 at $0.15 and Gemini Embedding 2 at $0.20 for text.
• Mistral costs more for less context: mistral-embed is $0.10 per 1M with an 8,000-token context, against $0.06 and 32,000 tokens on voyage-4.
• Free allowance covers a pilot: 200M free tokens per model embeds a mid-size corpus before paying, and querying on voyage-4-lite keeps the high-volume side at $0.02.

Reviewer’s Notes: I recommend starting on voyage-4 for both documents and queries. Once the index is stable, move the corpus to voyage-4-large through the Batch API, keep queries on voyage-4-lite, and add $100 of prepaid credits only if 2,000 requests per minute is not enough.

My Experience With Voyage AI

Signing Up and Getting an API Key

Creating an account generated a default organization and a first API key. I created a key in the Voyage dashboard under Organization > API keys and exported it as VOYAGE_API_KEY.

Voyage keeps billing at the organization level and keys at the project level, so I put development and production in separate projects:

• Organizations: billing, budget limits and data controls.
• Projects: their own API keys, rate limits and budget limits.
• Roles: Admin, Owner (project) and Member.

I added a payment method to reach Tier 1 at 2,000 requests per minute. Then, as Admin, I switched the Terms of Service toggle from Opted In to Opted Out, so my API data is not stored or used for training (zero-day retention).

That switch cannot be reversed from the dashboard. Re-opting in takes an email to Voyage’s legal team.

Embedding My First Documents

I installed the Python package with pip install voyageai (the TypeScript library is on npm as voyageai) and called the embed function with model voyage-4. I set input_type to document for the corpus and query for searches, and Voyage prepends a short retrieval prompt for each.

Each embed call takes up to 1,000 texts, with a per-call token cap that shrinks as the model grows:

• voyage-4-lite: 1M tokens per call.
• voyage-4: 320K tokens per call.
• voyage-4-large: 120K tokens per call.

Truncation is on by default, so over-length text is cut silently unless truncation is set to False. I previewed token counts with Client.tokenize and passed the model name, since without it the client loads the old tokenizer.

OpenAI embeddings follow the same embed-then-search pattern, so swapping clients in LangChain or LlamaIndex is a one-class change; both list Voyage integrations. The corpus still has to be re-embedded, because OpenAI and Voyage vectors live in different spaces.

Shrinking the Vectors

I set output_dimension to 512 and output_dtype to int8; uint8, binary and ubinary are the other compressed options. The models are Matryoshka-trained, so the first 256, 512 or 1024 values of a 2048-dimension vector work on their own.

int8 vectors take 4x less storage than float32, and binary vectors take 32x less. At 512 dimensions in int8, each vector is a sixteenth of the size of a 2048-dimension float32 vector.

Reranking the Results

I passed the query and the candidate documents to rerank-3, with top_k set to cap how many results came back. One call takes up to 1,000 documents, a query of up to 8,000 tokens and 600K total tokens.

Running a Batch Job

I uploaded my input file through the Files API and created a batch against the embeddings endpoint with the 12-hour completion window. A batch takes up to 100K inputs, and each embeddings request inside it still caps at 1,000 texts.

Reviewer’s Notes: Rate limits are the first wall to plan for: Tier 1’s 2,000 requests per minute comes with a token ceiling of only 3M per minute on voyage-4-large, and each project can carry its own lower limits. Make the opt-out a one-time Admin decision before any production data reaches the API.

The Voyage 4 Family: One Embedding Space, Four Model Sizes

voyage-4-large, voyage-4, voyage-4-lite and the open-weight voyage-4-nano, all launched on January 15, 2026, map text into one shared embedding space, so their vectors are interchangeable. All four take 32,000-token inputs and output 2048, 1024 (default), 512 or 256 dimensions.

• voyage-4-large: mixture-of-experts architecture at $0.12 per 1M tokens; Voyage says serving costs 40% less than comparable dense models.
• voyage-4: dense, $0.06 per 1M tokens, and my default for general text.
• voyage-4-lite: dense, $0.02 per 1M tokens, the pick for high-volume live queries.
• voyage-4-nano: dense, about 0.3B parameters (180M non-embedding plus 160M embedding), 2048 default dimensions, Apache 2.0 on Hugging Face with GGUF builds for Ollama and llama.cpp.

Voyage positions voyage-4-large for best quality, voyage-4 as the balance and voyage-4-lite for latency and cost. The shared space makes asymmetric retrieval cheap: embed documents once with voyage-4-large, embed queries with voyage-4-lite, and upgrade the query model later without re-indexing. voyage-4-nano runs local development offline against the same space as hosted production.

OpenAI’s text-embedding-3-small and text-embedding-3-large use separate spaces, so switching between them means re-embedding. Voyage’s compatibility does not reach backward: voyage-3.x vectors are not compatible with voyage-4, so an upgrade means re-embedding the corpus.

voyage-3-large, voyage-3.5 and voyage-3.5-lite went Legacy on January 15, 2026. Legacy models keep security and bug fixes, with at least 12 months notice before deprecation (6 months for rerankers).

voyage-code-4 is not stated to share the voyage-4 space, so I keep code vectors in a separate index from text vectors. Voyage says the voyage-4 series outperforms its own domain models, which turns voyage-law-2 and voyage-finance-2 into niche picks for existing indexes.

Rerankers and Contextualized Chunk Embeddings

rerank-3 and rerank-3-lite, released September 30, 2026, score a query against each candidate document with 32,000-token context and up to 1,000 documents per call. Prices stayed at $0.05 and $0.02 per 1M tokens, unchanged from the rerank-2.5 generation. Voyage pitches rerank-3 for highest accuracy and rerank-3-lite for latency, and the lite model gets double the Tier 1 throughput (4M tokens per minute versus 2M).

• Billing formula: query tokens x number of documents, plus the sum of document tokens.
• Long queries multiply cost: a 1,000-token query against 100 documents bills 100,000 query tokens before any document text.
• Request estimate: about $0.0025 for a 100-document rerank-3 request at 500 tokens each.
• Lifecycle: rerank-2.5 went Legacy on September 29, 2026.

Rerankers sit after any vector store, such as Pinecone or Qdrant: the store returns candidates and rerank-3 orders them. Cohere sells Rerank 5 with the same 32,000-token context, so Voyage is not the only one-vendor embed-plus-rerank stack.

voyage-context-4

voyage-context-4 (GA July 1, 2026) embeds each chunk with awareness of the full document, at $0.12 per 1M tokens, down from $0.18 for voyage-context-3. It runs on a mixture-of-experts backbone and is a drop-in replacement for voyage-context-3. It takes 32,000 tokens per chunk and 120,000 tokens of total context, so a chunk that says “the company” still carries the company name from elsewhere in the document.

• Auto-chunking: enable_auto_chunking splits documents using chunk_size and chunk_overlap.
• Pre-chunked input: lists of chunks keep your own boundaries, and overlapping chunks are supported natively.
• Queries: input_type set to query.

Code, Multimodal and Domain Models

voyage-code-4, released August 13, 2026, targets code search and coding agents at $0.12 per 1M tokens, a third below voyage-code-3’s $0.18. Voyage aims it at agentic retrieval, such as finding the files a fix touches from an issue description. It outputs 2048, 1024, 512 or 256 dimensions with int8 or binary quantization, sells through both the Voyage API and Atlas, and voyage-code-3 went Legacy the same day.

voyage-multimodal-3.5

voyage-multimodal-3.5 embeds interleaved text, images and video in one vector, which covers PDF screenshots, slides, tables and figures.

• Context: 32,000 tokens.
• Images: up to 16M pixels or 20 MB; every 560 pixels count as one token.
• Video: up to 20 MB.
• Price: $0.12 per 1M text tokens plus $0.60 per billion pixels.

Google’s Gemini Embedding 2 also embeds text, images, video, audio and PDFs in one space, at $0.20 per 1M text tokens, so it wins when audio is part of the corpus.

Legal and Finance Models

voyage-law-2 (16,000-token context) and voyage-finance-2 output 1024 dimensions only and cost $0.12 per 1M tokens with 50M free tokens each. Both went Legacy on January 15, 2026, and Voyage says the voyage-4 series now outperforms them.

Harvey, the legal AI company, built custom legal embeddings with Voyage starting from voyage-law-2. That custom-model route runs through Voyage’s sales team, not the self-serve dashboard.

Voyage AI Inside MongoDB Atlas

MongoDB acquired Voyage AI in February 2025, and Voyage models now sell two ways: the original Voyage API and the Atlas Embedding and Reranking API, a serverless, database-agnostic endpoint launched in public preview on January 15, 2026, with billing and monitoring inside the Atlas account.

Automated Embedding in MongoDB Vector Search embeds the data inside the database:

• Setup: an index field of type autoEmbed names the model; MongoDB embeds on initial sync, inserts, updates and query text.
• Where it runs: Atlas and self-managed Community Edition (mongot; MongoDB 8.2+ on the Kubernetes operator).
• Price: same list prices, 200M free tokens per model, one-time per organization; US or EU-scoped inference costs 1.1x.
• Free-tier ceiling: M0 clusters without a payment method get 3 requests per minute for auto-embedding queries; paid clusters get 2,000.
• Length handling: indexed text over 32,000 tokens is silently truncated, while over-length query text fails.
• Data use: MongoDB says this data is not used for training.

MongoDB’s launch note says Automated Embedding became generally available on Atlas on August 12, 2026, but its docs page still carries a preview label, so I treat the self-managed path as early.

Deploying in Your Own Cloud

• AWS Marketplace: model packages run on SageMaker inside your VPC, billed hourly for software plus instance; voyage-multilingual-2 costs $5.71 an hour plus about $1.41 an hour for one ml.g5.xlarge.
• Azure: managed applications run in your VNet at $5 an hour plus the VM price.
• Custom: on-prem deployments and model licensing go through sales.

Voyage lists SOC 2 and HIPAA compliance. You do not need MongoDB to use it: Voyage keys plug into Weaviate, Milvus or LanceDB, all of which support Voyage embeddings.

How Does Voyage AI Compare to Competitors?

At $0.02 to $0.12 per 1M tokens with 32,000-token inputs and a reranker on the same key, Voyage AI matches or undercuts every hosted embedding API with a published per-token price in this comparison.

Provider	Model	Price per 1M tokens	Open weights
Voyage AI	voyage-4-large / voyage-4-lite	$0.12 / $0.02	Yes (voyage-4-nano, Apache 2.0)
OpenAI	text-embedding-3-large / -small	$0.13 / $0.02	No
Google	gemini-embedding-001 / Gemini Embedding 2	$0.15 / $0.20 (text)	No
Cohere	Embed 5	No public per-token price	No
Mistral	mistral-embed	$0.10	No
Jina (Elastic)	jina-embeddings-v5-text-small	No public per-token price	Non-commercial license
Qwen	Qwen3-Embedding-8B	Self-hosted	Yes (Apache 2.0)

• OpenAI text-embedding-3: wins for teams already on the OpenAI API. Small matches voyage-4-lite at $0.02 and large costs $0.13, but OpenAI has no reranker (Voyage adds rerank-3 at $0.05 on the same key), and moving between small and large means re-embedding.
• Google Gemini embeddings: wins on multimodal breadth, since Gemini Embedding 2 covers text, images, video, audio and PDFs in one space. Text costs $0.15 (001) and $0.20 (Embedding 2) per 1M tokens, half price on batch. voyage-multimodal-3.5 handles text, images and video but not audio.
• Cohere Embed and Rerank: wins on input length, with Embed 5 taking 128,000 tokens and dimensions from 256 to 2048, plus Rerank 5 as its own reranker. Voyage’s 32,000-token embed limit is a quarter of that, but Cohere publishes no per-token Embed price.
• Jina embeddings v5 (now sold through Elastic): wins for Elasticsearch shops. v5-text-small has 677M parameters and a 32,768-token context, but the open weights are non-commercial without an Elastic license.
• Mistral embed: $0.10 per 1M with an 8,000-token context, two-thirds more than voyage-4 for a quarter of its input length, so it makes sense only if you already run on Mistral.
• Open-weight Qwen3-Embedding-8B and BGE-M3: win when data cannot leave your servers and you have GPUs. Qwen3 is Apache 2.0 with 8B parameters, a 32,000-token context and up to 4,096 dimensions; BGE-M3 is MIT with 8,192 tokens and 100+ languages.

voyage-4-nano is the lighter open option at about 0.3B parameters, and it sits in the same embedding space as Voyage’s hosted models, so a self-hosted prototype can move to the API without re-indexing.

How I Reviewed Voyage AI

I set up a Voyage organization and project on the Python client, created API keys, added a payment method and switched the data-training toggle to Opted Out. I embedded documents and queries with voyage-4-large, voyage-4 and voyage-4-lite, checked token counts with Client.tokenize, changed output dimensions and data types, reranked candidates with rerank-3 and ran a Batch API job through the Files API.

I weighed pricing and free tokens, model range, rate-limit tiers, data controls, the MongoDB Atlas route and the voyage-3.x migration cost. For rivals, I compared published prices and context limits from OpenAI, Google, Cohere, Mistral and Jina, plus the open-weight Qwen3-Embedding-8B and BGE-M3.

Prices current as of November 2026.

Voyage AI Review: Should You Use Voyage AI for Embeddings?

Yes, if you want cheap, long-context embeddings plus a reranker from one API, especially if you already run MongoDB. General models cost $0.02 to $0.12 per 1M tokens with 200M free tokens per model, rerank-3 adds candidate scoring at $0.05, and every voyage-4 model takes 32,000-token inputs.

The shared voyage-4 space is the reason to pick Voyage over OpenAI at the same $0.02 entry price: index once with voyage-4-large, query with voyage-4-lite, and change the query model later without rebuilding the index.

Skip Voyage AI if you need a hosted search product rather than model endpoints, if self-serve fine-tuning is a requirement, or if procurement demands contractual uptime credits: the 99.5% monthly figure is a Service Level Objective backed by commercially reasonable efforts. Pick Cohere Embed 5 for 128,000-token inputs and Gemini Embedding 2 for audio in the same space. On MongoDB, Automated Embedding embeds on insert and update, so no separate embedding pipeline is needed.

Switch the data setting to Opted Out before sending private documents. Then embed a sample of your own corpus with voyage-4 inside the free allowance at Voyage AI, and compare it against your current provider on your own queries.

FAQ

Is Voyage AI free?

Voyage AI gives every account 200M free tokens per current model, then bills per token. voyage-law-2, voyage-finance-2 and voyage-code-2 get 50M free tokens each, and voyage-multimodal-3.5 adds 150B free pixels. Free tokens do not apply to Batch API usage or to older voyage-3 generation models.

Is Voyage AI owned by MongoDB?

Yes. MongoDB acquired Voyage AI in February 2025, and the product is now branded Voyage AI by MongoDB. The original Voyage API runs alongside the Atlas Embedding and Reranking API, and neither needs a MongoDB database. Only Automated Embedding requires MongoDB Vector Search.

Does Voyage AI train on my data?

By default, Voyage AI may store and use API data for training. An org Admin with a payment method on file can switch the setting to Opted Out for zero-day retention. The dashboard cannot reverse that choice; re-opting in takes an email to Voyage’s legal team. MongoDB says data processed through Automated Embedding is not used for training.

Can I mix Voyage 4 models for documents and queries?

Yes. voyage-4-large, voyage-4, voyage-4-lite and voyage-4-nano share one embedding space, so you can index documents with large and query with lite at the same output dimension. Older voyage-3.x vectors are not compatible, so that migration means re-embedding the corpus.

What are Voyage AI’s rate limits?

Tier 1 requires a payment method and allows 2,000 requests per minute on every model. Token limits vary by model: 16M per minute on voyage-4-lite, 8M on voyage-4 and voyage-code-4, 3M on voyage-4-large and voyage-context-4, and 2M on rerank-3. Tier 2 doubles those limits after $100 paid, Tier 3 triples them after $1,000, and $100 of prepaid credits jumps straight to Tier 2.