Voyage AI Embeddings: Vector Dimensions, Retrieval Quality, Context Length, and Pricing
Voyage AI builds embedding and reranking models used as the retrieval layer in RAG pipelines, sold as a hosted API and, since February 2025, owned by MongoDB. The acquisition folded Voyage's models into Atlas Vector Search as an automated embedding option, while the standalone API continues to serve any vector store — Pinecone, Qdrant, pgvector, Milvus, Weaviate, Elasticsearch, OpenSearch, and others. In January 2026 Voyage replaced its voyage-3 generation with voyage-4, a model family built on a shared embedding space and, for the flagship model, a mixture-of-experts architecture. This article covers the current model lineup, how vector dimensions and quantization are controlled, what Voyage's own retrieval benchmarks report against OpenAI and Cohere, and where the pricing and context-length trade-offs actually bind in production.
········
VOYAGE 4 MODEL LINEUP: CONTEXT LENGTH, DIMENSIONS, AND PRICING.
Which model fits a given retrieval workload and what it costs per million tokens processed.
Voyage's current catalog centers on four voyage-4 models that share one embedding space, plus a contextualized-chunking variant and a code-specialized model released alongside them. Two domain models (law, finance) remain on the older voyage-2 generation and have not been updated to voyage-4 as of this writing. All voyage-4-family models support a 32,000-token context window and Matryoshka Representation Learning (MRL), which lets a single embedding be truncated to smaller dimensions without retraining.
........
Model | Context length | Output dimensions | Price / 1M tokens | Monthly free tokens | Primary use case |
voyage-4-large | 32,000 tokens | 256 / 512 / 1024 / 2048 | $0.12 | 200M | Flagship, MoE architecture, best reported general-purpose & multilingual retrieval |
voyage-4 | 32,000 tokens | 256 / 512 / 1024 / 2048 | $0.06 | 200M | Balanced quality/cost, default recommendation |
voyage-4-lite | 32,000 tokens | 256 / 512 / 1024 / 2048 | $0.02 | 200M | Latency- and cost-optimized; query-side use in shared embedding space |
voyage-4-nano | 32,000 tokens | 256 / 512 / 1024 / 2048 | Open-weight (Apache 2.0) | Self-hosted, no API cost | On-device or self-hosted inference |
voyage-context-4 | 32,000 tokens per pass; longer documents auto-split | 256 / 512 / 1024 / 2048 | $0.12 | 200M | Contextualized chunk embeddings with built-in auto-chunking |
voyage-code-4 | 32,000 tokens | 256 / 512 / 1024 / 2048 | $0.12 | 200M | Code retrieval and coding-agent context |
voyage-law-2 (gen-2) | 16,000 tokens | 1024 | $0.12 | 50M | Legal-domain retrieval |
voyage-finance-2 (gen-2) | 32,000 tokens | 1024 | $0.12 | 50M | Financial-domain retrieval |
........
Every account receives its monthly free-token allowance before per-token billing applies; batch API usage is discounted further on the standard voyage-4 models. The domain models still carry gen-2 pricing and a smaller 50M free allowance, reflecting their older release cycle relative to the voyage-4 family.
········
DIMENSIONS, QUANTIZATION, AND THE SHARED EMBEDDING SPACE.
How output size and numeric precision are chosen independently of retrieval quality, and what changed structurally in voyage-4.
Each voyage-4 model is trained with Matryoshka Representation Learning, meaning the first 256, 512, or 1024 values of the full 2048-dimensional vector are themselves valid, independently useful embeddings. Truncating dimensions after the fact — without recomputing anything — trades some retrieval quality for smaller index storage and faster distance computation, and the degradation is designed to be gradual rather than a cliff. Separately from dimension count, each embedding can be output at three numeric precisions: 32-bit float, signed or unsigned 8-bit integer, or binary (1-bit per dimension). For the prior voyage-3-large generation, Voyage AI reported int8 quantization at 1024 dimensions costing only 0.31% in quality versus full float precision while using roughly 8x less storage, and binary embeddings at 512 dimensions beating OpenAI's float text-embedding-3-large (3072 dimensions) by 1.16% on their evaluation set while using around 200x less storage — a vendor-reported figure specific to that model generation, not independently reproduced here.
The structural change in voyage-4 is a shared embedding space across all four general-purpose models (voyage-4-large, voyage-4, voyage-4-lite, voyage-4-nano). Previous Voyage generations required the same model at indexing and query time. With voyage-4, a corpus can be indexed once with voyage-4-large for maximum quality, then queried with voyage-4-lite or the open-weight voyage-4-nano to cut per-query latency and cost — an asymmetric retrieval pattern that avoids full reindexing when the query-side cost profile needs to change. Voyage AI reports that quality holds or improves under this asymmetric setup across the domains it tested (medical, code, web, finance, documentation, legal, conversations, long documents), though this too is a vendor-reported result rather than a third-party benchmark.
········
RETRIEVAL QUALITY: VOYAGE'S BENCHMARKS AGAINST OPENAI, COHERE, AND GEMINI.
What Voyage AI's own published evaluation reports, and where the reranking layer fits alongside the embedding model.
Voyage AI's benchmark for voyage-4-large uses RTEB, a 29-dataset retrieval suite spanning medical, code, web, finance, documentation, legal, conversational, and long-document domains. These are vendor-published numbers; Voyage AI has not published raw per-dataset results validated by an independent third party at the time of writing, and other aggregator sites report a different margin (around 8.2% on NDCG@10) against Cohere's embed-v4 than Voyage's own RTEB figure below — the two use different evaluation sets, so they are not directly comparable.
........
Model compared | RTEB score vs voyage-4-large | Source |
voyage-4 | −1.87% | Voyage AI, 29-dataset RTEB average |
voyage-4-lite | −4.80% | Voyage AI, 29-dataset RTEB average |
Cohere embed-v4 | −3.87% | Voyage AI, 29-dataset RTEB average |
Gemini embedding-001 | −8.20% | Voyage AI, 29-dataset RTEB average |
OpenAI text-embedding-3-large | −14.05% | Voyage AI, 29-dataset RTEB average |
........
Voyage's reranking models are sold as a separate pass applied after initial vector or hybrid retrieval, not a replacement for the embedding step. The current generation, rerank-2.5 and rerank-2.5-lite, supports a 32,000-token context window — Voyage states this is 8x Cohere Rerank v3.5's context and double the prior rerank-2 generation — and adds instruction-following, letting a caller bias reranking toward a document section, type, or disambiguation rule at query time rather than through prompt engineering upstream. Voyage AI reports rerank-2.5 improving retrieval accuracy by 7.94% over Cohere Rerank v3.5 on standard datasets and by 12.70% on the MAIR benchmark, with rerank-2.5-lite close behind at 7.16% and 10.36% respectively; pricing is token-based and existing rerank-2 customers get the longer context at the same per-token rate. As with the embedding benchmarks, these are Voyage-reported comparisons.
········
CONTEXT, CHUNKING, AND COST IN A PRODUCTION RAG PIPELINE.
Where the context-length and dimension choices above actually show up as reindexing cost, lock-in, and deployment constraints.
The 32,000-token context ceiling on voyage-4 models applies per embedding call, which matters less for short chunks than for whole-document embedding strategies. voyage-context-4 changes the unit of work: instead of embedding pre-split chunks independently and losing whatever surrounding context explained a pronoun, a defined term, or a cross-reference, it processes the full document in one pass and emits one context-aware vector per chunk, auto-splitting anything beyond the 32K window transparently. Voyage AI reports this improves chunk-level retrieval by 2.08% and document-level retrieval by 1.4% over its predecessor across 39 datasets, and that voyage-context-4 even beats voyage-4-large as a plain single-vector embedder by 0.45% — again vendor-reported, and worth validating against a held-out sample of the actual corpus before committing an indexing pipeline to it, since chunking-sensitive gains vary heavily with document structure.
Cost scales with tokens embedded, not documents, so reindexing a large corpus after a model or dimension change is the dominant cost event, not the steady-state query load — steady-state querying is a small volume per request against a 200M-token monthly free allowance. This is precisely what the voyage-4 shared embedding space is designed to avoid: swapping the query-side model (large to lite or nano) needs no reindex, but changing the index-side model, moving to voyage-context-4's chunking model, or increasing dimensions all require re-embedding the full corpus. Vendor lock-in is structural, not contractual: embeddings from Voyage's space are not interchangeable with OpenAI's, Cohere's, or any other provider's vectors, so a switch later means a full reindex regardless of vector database choice. Availability is API-only for every model except voyage-4-nano, which is Apache 2.0-licensed and self-hostable, making it the only option for teams that cannot send document content to a third-party API. Decision rule: default to voyage-4 for general RAG indexing and query with voyage-4-lite once query volume grows, since the shared embedding space makes that split free; use voyage-4-large only when RTEB-style retrieval accuracy is the binding constraint and the corpus is stable enough to amortize a full-precision reindex; switch to voyage-context-4 specifically when chunks are pulled from long, cross-referential documents (contracts, transcripts, technical manuals) where isolated-chunk embeddings demonstrably lose meaning; and choose voyage-4-nano over the API only when self-hosting or data residency rules out sending content to any external embedding endpoint.
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
