Leading Embedding Model Pricing & Dimension Sizing Matrix (2026)
Verified dimensional requirements and cost per 1M tokens across major embedding providers.
| Model Name | Dimensions | Price / 1M Tokens | RAM / 100k Vectors | Ideal Use Case | Official Docs |
|---|---|---|---|---|---|
| text-embedding-3-small (OpenAI) | 1,536 | $0.020 | ~614 MB | Production RAG & Search | OpenAI Docs ↗ |
| text-embedding-3-large (OpenAI) | 3,072 | $0.130 | ~1.23 GB | High-Precision Legal/Medical | OpenAI Docs ↗ |
| Embed English v3.0 (Cohere) | 1,024 | $0.100 | ~410 MB | Enterprise Multi-Stage RAG | Cohere Docs ↗ |
| BGE-M3 (Open-Source) | 1,024 | $0.000 (Compute) | ~410 MB | Air-Gapped & Local Search | Hugging Face ↗ |
RAG Token Economics & Vector Database Sizing
Retrieval-Augmented Generation (RAG) is the dominant architecture for grounding foundation models on enterprise knowledge. While embedding tokens are relatively inexpensive compared to generation tokens, the operational expense of RAG systems is dominated by the vector database running in RAM to power low-latency approximate nearest neighbor (ANN) search.
Vector Dimension Memory Breakdown
Every vector embedding consists of floating-point numbers. In standard single-precision (Float32), each dimension occupies exactly 4 bytes:
- 768 dimensions = 3,072 bytes per vector
- 1536 dimensions = 6,144 bytes per vector
- 3072 dimensions = 12,288 bytes per vector
When utilizing graph-based indexes like HNSW, the index must store multi-layer graph adjacency lists in RAM to achieve sub-10ms search times, requiring roughly 1.5× the raw vector size in available memory.
Frequently Asked Questions
How do you calculate RAG embedding costs?
Multiply document count by chunks per document to get total chunks, then multiply by chunk size to get total tokens. Divide by 1,000,000 and multiply by the provider's embedding rate (e.g., $0.02/1M for OpenAI small).
How much RAM does a vector database require?
For HNSW (Hierarchical Navigable Small World) indexes used in Pinecone, Qdrant, and pgvector, each Float32 dimension takes 4 bytes. An additional ~1.5x multiplier is needed for graph links and index metadata in memory.
What chunk size is recommended for RAG?
Between 300 and 800 tokens with a 10% to 15% overlap is standard for general retrieval. Smaller chunks preserve precision; larger chunks retain broad semantic context.