Voyage AI Pricing & Embedding Rates
Voyage AI builds state-of-the-art vector embedding and reranking models designed specifically for RAG retrieval. Compare per-token embedding rates, domain-specific models, and retrieval accuracy.
Enterprise Vector Embedding & RAG Cost Simulator
Simulate monthly embedding costs for knowledge bases, codebases, and legal archives.
Production Pricing & Model Tier Rates
Verified rates, context capacities, typical inference latency, and recommended enterprise workloads.
| Model / Tier | Input Rate (1M) | Output Rate (1M) | Context Window | Latency / Throughput | Recommended Workload |
|---|---|---|---|---|---|
| voyage-3 (General) | $0.120 / 1M | N/A | 32,000 | Sub-150ms | #1 general English & multilingual retrieval benchmark |
| voyage-3-lite | $0.050 / 1M | N/A | 32,000 | Sub-80ms | High-volume search, latency-sensitive consumer apps |
| voyage-code-3 | $0.120 / 1M | N/A | 32,000 | Sub-150ms | Code search, repo indexing, function signature retrieval |
| voyage-finance-2 | $0.120 / 1M | N/A | 32,000 | Sub-150ms | Financial 10-K filings, earnings transcripts, SEC docs |
| voyage-law-2 | $0.120 / 1M | N/A | 32,000 | Sub-150ms | Legal contracts, statutory case law, compliance discovery |
| rerank-2 | $0.050 / 1k queries | N/A | 16,000 | Sub-100ms | Second-stage cross-encoder re-ranking for maximum precision |
⚡ Architectural Deep Dive & Unit Economics Analysis
Voyage AI was founded by Stanford AI professor Percy Liang to solve the fundamental Achilles' heel of enterprise RAG systems: poor retrieval quality. Standard generalist embedding models (such as OpenAI text-embedding-ada-002 or text-embedding-3) frequently fail to rank subtle technical domain clauses in the top-k results, causing LLMs to hallucinate.
Voyage-3 consistently outperforms OpenAI and Cohere on the Massive Text Embedding Benchmark (MTEB), offering specialized models trained specifically for source code (voyage-code-3), financial filings (voyage-finance-2), and law (voyage-law-2). Crucially, Voyage supports an expansive 32,000-token context window per chunk—4x larger than OpenAI's 8,191-token ceiling—allowing whole document chapters to be embedded without breaking semantic coherence.
Production Workloads & Monthly Cost Modeling
Evaluating practical unit economics across real-world application architectures.
Enterprise Codebase Semantic Search (Cursor/IDE)
Embedding 50M tokens of company repositories using voyage-code-3. Unlocks precision code snippet lookup for $6.00 total.
Financial Audit & SEC 10-K RAG Assistant
Analyzing 20M tokens of annual reports using voyage-finance-2. Outperforms OpenAI by 18% on tabular financial data retrieval.
Two-Stage Retrieval (Embed + Rerank)
Retrieving top 50 candidates with voyage-3-lite, then refining top 5 with rerank-2. Delivers 95%+ precision at minimal compute cost.
import os
import voyageai
# Initialize Voyage AI official client
vo = voyageai.Client(api_key=os.environ.get("VOYAGE_API_KEY"))
def embed_documents(texts: list[str], model: str = "voyage-3"):
"""Embed documents with 32,000 token context window support."""
result = vo.embed(
texts=texts,
model=model,
input_type="document" # 'document' for storage, 'query' for search
)
return result.embeddings
def rerank_search_results(query: str, documents: list[str], top_k: int = 5):
"""Re-rank candidates using cross-encoder accuracy."""
reranking = vo.rerank(
query=query,
documents=documents,
model="rerank-2",
top_k=top_k
)
return [r.document for r in reranking.results]Operational Gotchas & Failure Modes
Engineering realities, throttling thresholds, and hidden surcharge traps to monitor.
Related Models, Showdowns & Benchmarks
Cross-reference provider rates against hosted models, direct head-to-head battles, and calculators.
Voyage AI vs OpenAI Embeddings
Compare MTEB retrieval scores and cost benchmarks.
Cheapest Embedding APIs of 2026
Compare Voyage, OpenAI, Cohere, and open-source embedding models.
Pinecone vs Qdrant Showdown
Compare vector databases for storing Voyage embeddings.
Chat Context Bloat Calculator
Calculate how vector retrieval cuts LLM prompt token costs.
Frequently Asked Questions
Is Voyage AI better than OpenAI embeddings?
Yes. Voyage-3 outperforms OpenAI text-embedding-3-large on the official MTEB retrieval benchmark, offering higher retrieval recall, specialized domain models, and a 32,000 token context window.
What is the token pricing for Voyage AI?
voyage-3 is priced at $0.12 per million tokens. voyage-3-lite is priced at $0.05 per million tokens. Domain-specific models (code, finance, law) are $0.12 per million tokens.
How large is the context window for Voyage embeddings?
Voyage models support up to 32,000 tokens per embedding request, which is roughly 24,000 words (4x larger than OpenAI's 8,191-token limit).
What is the input_type parameter in Voyage AI?
Voyage trains asymmetric embedding models. You specify input_type='document' when creating embeddings to save into your database, and input_type='query' when generating embeddings for user search questions.
Does Voyage AI have a free tier?
Yes! Voyage AI provides 50 million free tokens upon account registration, allowing you to embed entire knowledge bases without adding a credit card.
What is Voyage rerank-2?
rerank-2 is a cross-encoder model that takes a search query and a list of candidate documents and re-orders them by semantic relevance, drastically improving RAG answer accuracy.
Voyage AI Domain-Specific Embeddings: Recall Economics & Reranking Models
Voyage AI is the industry standard for high-accuracy embedding models, frequently topping MTEB retrieval leaderboards. Unlike generic hyperscaler embeddings, Voyage trains specialized foundation models for specific domains: voyage-code-3 for source code repositories, voyage-finance-2 for financial SEC filings, and voyage-law-2 for legal discovery. At $0.12 per million tokens ($0.02/M on voyage-3-lite), Voyage models deliver up to 15% higher Top-10 recall than OpenAI text-embedding-3-small, preventing costly downstream LLM hallucination repair turns.
Voyage AI Model Benchmark & Rate Card
| Model Name | Dimensions | Context Window | Cost / 1M Tokens | Primary Specialty |
|---|---|---|---|---|
| voyage-3-lite | 512 | 32,000 tokens | $0.020 | Ultra-low RAM vector indexing |
| voyage-3 | 1,024 | 32,000 tokens | $0.120 | State-of-the-art general retrieval |
| voyage-code-3 | 1,024 | 32,000 tokens | $0.120 | Codebase AST & documentation search |
| rerank-2 | Cross-Encoder | 16,000 tokens | $0.050 / 1k queries | Precision cross-encoder reranking |
Voyage AI Codebase Vectorization Client (Python)
import voyageai
vo = voyageai.Client()
# Batch embed code chunks with 32k context support
texts = ["def calculate_tco(instances, hours): ...", "class DatabaseConnectionPool: ..."]
result = vo.embed(texts, model="voyage-code-3", input_type="document")
print(f"Generated {len(result.embeddings)} vectors of dimension {len(result.embeddings[0])}")Engineering Q&A: Voyage AI Embeddings
What is the benefit of voyage-3-lite's 512-dimension output?
In vector stores like Qdrant and Milvus, 512-dimension vectors consume 66% less memory than 1,536-dimension vectors. For a 10-million chunk enterprise index, this shrinks server RAM requirements from 62GB to under 21GB, saving over $300/month in database hosting.
Does Voyage support asymmetric query vs document embedding?
Yes. Always specify `input_type="query"` for search queries and `input_type="document"` for knowledge base chunks. Voyage applies specialized instruction prompts during encoding to maximize asymmetric cross-attention recall.