Embedding Model & Vector DB Cost FAQ
7 common questions · all prices and scores are reference estimates (compiled 2026-06-20)
Frequently asked questions (7)
Which embedding model is cheapest for RAG?
By per-1M-tokens price, Google text-embedding-005 ($0.006) is lowest, followed by OpenAI text-embedding-3-small, Voyage voyage-3.5-lite and Jina v3 (around $0.02). Self-hosted BGE-M3 has no API token cost, only GPU and infrastructure. But total cost also includes vector storage, which is driven by dimension, so a cheap model with a large dimension can cost more overall. Combine both in the calculator.
How is vector storage size calculated?
Vectors are stored as float32 numbers, so storage is embedding dimension × number of documents × 4 bytes. A single 1,024-dimension vector takes 4,096 bytes (4 KB), and ten million such vectors take about 38 GiB. A larger dimension such as 3,072 scales storage proportionally. That is why Matryoshka reduction or quantization is the key lever for lowering cost.
What is Matryoshka embedding reduction?
Models like OpenAI v3, Voyage, Cohere v4, Jina v3 and Gemini are trained so you can truncate a vector from the front to a shorter dimension with little quality loss. Reducing text-embedding-3-large from 3,072 to 256 dimensions cuts storage to one-twelfth, sharply lowering vector-DB cost at a modest accuracy trade-off you can validate on your own data.
Does a higher MTEB score always mean a better model?
Not necessarily. MTEB is an average across many tasks, so it does not guarantee performance on your specific domain or language. A model with slightly lower MTEB but a longer context, lower price or stronger multilingual support may fit you better. For Korean or multilingual RAG, validate multilingual models on your own eval set rather than trusting English MTEB alone.
Are the prices and scores exact?
No. Every price and MTEB score here is a reference estimate compiled from public sources and subject to change (compiled 2026-06-20). Providers update pricing and release new models frequently, so always confirm current rates in each provider console before committing. This tool is for rough planning and comparison, not for billing.
Which is cheaper: Pinecone, Qdrant or pgvector?
This tool models only the storage rate ($/GB·month), roughly ordered pgvector < Qdrant < Pinecone. Real costs vary widely with read/write units, RAM, replicas and minimums, and pgvector adds self-management overhead. Weigh managed convenience against operating cost when deciding, and treat the figures here as a starting estimate, not a quote.
Is there a separate query cost?
Yes. Every search embeds the query text, which incurs token cost. Queries are usually just tens of tokens, so this is often smaller than the document indexing cost. The calculator estimates monthly query-embedding cost as monthly queries × average query tokens and adds it to the monthly total. Reranking and LLM generation costs are separate and not included here.