Match each embedding provider to a defining feature or pricing fact

Each major embedding vendor has a defining selling point. Knowing them is faster than reading every model card. OpenAI text-embedding-3-small is the cost-optimized default: USD 0.02 per million tokens, 1536 dimensions, Matryoshka-truncatable (can be shortened to fewer dimensions). Its sibling text-embedding-3-large is the premium quality option at USD 0.13 per million. Voyage voyage-code-3 is the code specialist, distinct from Voyage's general voyage-3 model. A code-heavy training mix gives it about a 14 percent average edge over OpenAI's large model on code retrieval benchmarks. BGE-M3 is the open-weights multilingual leader from BAAI. It handles 100+ languages and outputs three kinds of representations (dense, sparse, multi-vector) from a single forward pass. Cohere embed-v3 and v4 are tightly paired with the rerank-3.5 cross-encoder. That makes them the popular choice for production RAG stacks that need both a first-stage embedder and a reranker (a model that reorders search results). Nomic Embed v1.5 is the open-weights self-host pick. It is Matryoshka-trained, runs on a single GPU, and sends no data to an outside API.

Each vendor in the lineup has a distinct positioning that drives selection. **OpenAI text-embedding-3-small** is the cost-optimized default at $0.02 per million tokens. It offers 1536 dimensions, Matryoshka truncation, hosting, and low operational overhead. It is the 'just pick this' answer for general English RAG when you do not want to think about model selection. **Voyage voyage-code-3** is the code specialist. A code-heavy training mix gives it an average 13.8 percent edge over OpenAI text-embedding-3-large across 32 code retrieval datasets, per Voyage. **BGE-M3** is the open-weights multilingual leader from BAAI. It covers 100+ languages with a shared embedding space, plus hybrid output (dense, sparse, and multi-vector from one call). It is self-hostable on a single H100. **Cohere embed-v3 and v4** ship tightly paired with rerank-3.5, a cross-encoder reranker from the same vendor. The pairing is the value: picking embedder and reranker from one vendor simplifies the stack and improves end to end quality. Nomic Embed v1.5 is the on-prem open-weights pick. It is Matryoshka-trained, strong on MTEB, and designed for the 'no data sent to outside APIs' constraint that enterprise and regulated deployments often have.