Voyage AI
Listed at https://voyageai.com
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | voyage-4-large: $0.12/M, voyage-4: $0.06/M, voyage-4-lite: $0.02/M, voyage-context-3: $0.18/M, voyage-code-3: $0.18/M, voyage-multimodal-3.5: $0.12/M | per million tokens |
| Output | （embedding API — no output tokens; rerank-2.5: $0.05/M tokens, rerank-2.5-lite: $0.02/M tokens） | per million tokens |
🤖 Supported Models (8)
✨ Pros
- ✓voyage-4 series surpasses OpenAI v3 Large by 14%, Cohere Embed v4 by 8%, Gemini Embedding 001 by 4% on RTEB benchmark
- ✓MoE architecture: voyage-4-large is the first production-grade MoE embedding model
- ✓32K context window (vs OpenAI's 8K, Cohere's 512)
- ✓Matryoshka flexible dimensions: 2048/1024/512/256 dims for vector DB cost optimization
- ✓200M free tokens on signup (embeddings + rerankers)
- ✓Domain-specific models: voyage-code-3, voyage-finance-2, voyage-law-2
- ✓Joined MongoDB team (March 2025) — backed by a major database vendor
⚠️ Cons
- ×API not OpenAI-compatible (different schema, requires input_type param)
- ×No chat completions endpoint (only embedding + rerank + multimodal)
- ×China access stability not officially documented
- ×No public affiliate program
- ×Focused on embedding/rerank, no general chat models
🎯 Best For
RAG systems, vector retrieval, document embeddings, code search, multimodal (text+image) retrieval; enterprise search apps needing high-accuracy embeddings
💰 Pricing & Plans
| Model / Endpoint | Price per 1M tokens | Free allowance | Best For |
|---|---|---|---|
| voyage-4-large | $0.12 | 200M tokens | Best general-purpose and multilingual retrieval quality |
| voyage-4 | $0.06 | 200M tokens | Balanced quality and cost; vectors interoperate across the 4 series |
| voyage-4-lite | $0.02 | 200M tokens | Lowest latency and cost for high-volume bulk indexing |
| voyage-code-4 | $0.12 | 200M tokens | Code retrieval and coding-agent use cases |
| voyage-context-4 | $0.12 | 200M tokens | Contextualized chunk embeddings for long documents |
| voyage-finance-2 / voyage-law-2 | $0.12 | 50M tokens | Domain-specific retrieval for finance and legal corpora |
| voyage-multimodal-3.5 | $0.12 text + $0.60 per 1B pixels | 200M text tokens + 150B pixels | Text, image and video embedding through one endpoint |
| rerank-3 | $0.05 | 200M tokens | Second-stage reranking of candidates returned by vector search |
| rerank-3-lite | $0.02 | 200M tokens | Cheaper reranking across very large candidate pools |
| Batch API (embeddings + rerank) | 33% discount on standard rates | Free tokens do not apply | Bulk corpus indexing with a 12-hour completion window |
| Files API storage | $0.05 per GB per month | Not applicable | Batch input and output files, 30-day retention |
| Legacy: voyage-context-3 / voyage-code-3 | $0.18 | No free tokens | Pinned pipelines still running the previous generation |
🔧 API & Developer Experience
- •REST API + OpenAPI spec: Three inference endpoints (/v1/embeddings, /v1/contextualizedembeddings, /v1/rerank) plus a public OpenAPI spec at github.com/voyage-ai/openapi for building custom clients
- •Official Python client + community SDKs: The voyageai package exposes embed(), contextualized_embed() and rerank(); community-maintained clients cover Ruby, Vercel, Go, TypeScript and C#
- •input_type parameter: Pass query or document so the encoder knows the retrieval role; Voyage's own docs recommend it for every RAG workload
- •Framework and vector database integrations: LangChain, LlamaIndex, Haystack, LanceDB, Weaviate, Pinecone, Qdrant, Milvus, Redis, pgai and Unstructured all ship first-class Voyage adapters
- •Rate limits and usage tiers: 2,000 RPM on every model family with per-model TPM ceilings from 2M (rerank-3, multimodal) to 16M (voyage-4-lite); limits graduate automatically as organization spend grows
- •Batch API and Files API: Bundle requests into JSONL for a 12-hour window at a 33% discount, with a Files service storing inputs and outputs for 30 days at $0.05 per GB per month
- •Self-hosted deployment options: Model packages on AWS Marketplace and an Azure Marketplace managed application run inference inside your own VPC
🔍 Contextualized Chunk Embeddings
voyage-context-4 is the one capability no general-purpose embedding API offers: contextualized chunk embeddings. Instead of vectorizing each chunk in isolation, it embeds chunks together with the surrounding document, so a vector carries both what the chunk says and how it functions in the whole piece, which is the leading failure mode in production RAG. Each chunk gets a 32,000-token context window, and with enable_auto_chunk=true total input reaches 120,000 tokens, so a long document is pushed once rather than re-sliced by hand. Dimensions are Matryoshka-trained, so 2048/1024/512/256 come from one pass, and int8 or binary quantization cuts vector storage by 4x to 32x.
🌐 China Access & Latency
Voyage AI runs inference on Google Cloud: api.voyageai.com resolves to a GCP address and returns via: 1.1 google with x-served-by: prod1, while the marketing site is served from Vercel on Singapore PoP routing. Neither is a mainland China edge, and Voyage publishes no China-specific endpoint, region or availability commitment in its docs, so treat direct calls from the Chinese mainland as proxy-dependent. Measured from an Alibaba Cloud Malaysia host on 29 October 2026, the docs site returned TTFB in 48-69ms, the marketing site 60-68ms, and the API endpoint 240-725ms unauthenticated; behind a proxy, Chinese users should expect roughly 400-900ms per embedding call, acceptable for batch corpus indexing but noticeable for interactive reranking. Two mitigations matter for China-facing teams: the Batch API absorbs index-building volume at a 33% discount without extra round trips, and the AWS and Azure Marketplace model packages let you run the models inside your own VPC in Hong Kong or Singapore, keeping document content on infrastructure you control. The published SLO is 99.5% monthly uptime for api.voyageai.com.