Voyage AI
Embedding and reranking API, now owned by MongoDB: Voyage 4 models with 32K context, code, finance and law variants, 200M free tokens
Overview
Voyage AI builds embedding and reranking models for search and RAG. MongoDB announced it was buying the company in February 2025 (Bloomberg reported $220 million), and the site now reads "Voyage AI by MongoDB", but the API is still sold on its own. The Voyage 4 series (January 2026) includes voyage-4-large, voyage-4 and voyage-4-lite, which share one embedding space so you can index with the large model and query with a cheaper one, plus voyage-4-nano, a small open-weight model under Apache 2.0. All take 32K tokens and output 256 to 2,048 dimensions, with int8 and binary options on the main models. Domain models cover code (voyage-code-4), finance and law, and voyage-multimodal-3.5 embeds images. Rerankers (rerank-3 and rerank-3-lite) round out the line. Co-founder Tengyu Ma is a Stanford professor. Also from MongoDB: MongoDB Atlas Vector Search, which can now call Voyage models directly.
The Verdict
Who Should Use Voyage AI?
Best For
- RAG teams whose bottleneck is retrieval quality
- Code search and coding agents (voyage-code-4)
- Legal and finance document search
- MongoDB Atlas users who want embeddings in the database
Not Ideal For
- Fully self-hosted stacks: only voyage-4-nano has open weights
- Inputs over 32K tokens per chunk (Cohere takes 128K)
- Buyers wary of a roadmap set by a database vendor
What's Great
- Led a 2026 hands-on RAG test: Recall@10 0.89 vs 0.86 for OpenAI 3-large
- 200M free tokens on each current model
- Shared embedding space across the Voyage 4 family
- Batch API at 33% off
- Used by Harvey, Replit, Snowflake, kapa and Continue
Watch Out For
- Headline wins are Voyage's own RTEB results; the independent gap was about 3 points
- The pricing page contradicts itself on free tokens for some older models
- Many generations to track; older v2 and v3 models carry no free tokens
- Very few public reviews (G2 shows none)
Pricing
View all features & details
Current Models
- voyage-4-large, voyage-4, voyage-4-lite: 32K, 1,024 dims by default, 256 to 2,048
- voyage-4-nano: open weights, Apache 2.0, about 0.3B parameters
- voyage-code-4: 32K, for code retrieval
- voyage-finance-2 (32K) and voyage-law-2 (16K): 1,024 dims
- voyage-multimodal-3.5: $0.12 per 1M text tokens plus $0.60 per 1B pixels
- voyage-context-4: contextualized chunk embeddings
Options
- Output types: float, int8, uint8, binary, ubinary (Voyage 4 family)
- Batch API: 33% discount, 12-hour window
- Files API storage $0.05 per GB per month
Where It Runs
- Voyage API (standalone)
- MongoDB Atlas Embedding and Reranking API (public preview in early 2026)
- AWS and Azure marketplaces
Real-World Usage
Customers
- Harvey, Replit, Snowflake, kapa, Continue (site logos)
- Tavily and TinyFish CEOs quoted at the Voyage 4 launch
Ownership
- MongoDB deal announced 2025-02-24; reported price $220 million
- Plan: keep the standalone API, add auto-embedding in Atlas, then multimodal models
- Atlas Embedding and Reranking API launched in preview in 2026
How It Compares
| Feature | Voyage 4 | OpenAI 3-large | Cohere Embed | Nomic Embed |
|---|---|---|---|---|
| Recall@10 (Techsy, 2026) | 0.89 (4-large) | 0.86 | 0.85 (v4.0) | Not tested |
| Max input | 32K tokens | 8,192 tokens | 128K tokens | 8,192 tokens (v1.5) |
| Price per 1M | $0.02 to $0.12 | $0.02 to $0.13 | Not published | Free to self-host |
| Free allowance | 200M tokens | None listed | Trial key | Open weights |
| Domain models | Code, finance, law | No | No | Code |
Sources: Techsy · Voyage docs · OpenAI pricing · Cohere models · Nomic card
Is this your tool? Add a badge