Dev Tools

Embeddings Cost Calculator

Compare cost across OpenAI, Voyage AI, and Cohere embedding APIs.

Quick Answer

text-embedding-3-small at $0.02/M is the cost-leader for most RAG workloads. text-embedding-3-large at $0.13/M trades 6.5x cost for ~5-10% recall gain. Voyage 3-large ($0.18/M) tops MTEB but costs the most. Embedding 1M docs at 500 tokens each costs $10 on small, $65 on large.

ModelDimsPer 1MInitialYear 1
text-embedding-3-small
OpenAI
1536$0.02$1.00$1.60
text-embedding-3-large
OpenAI
3072$0.13$6.50$10.40
text-embedding-ada-002
OpenAI · Legacy
1536$0.1$5.00$8.00
voyage-3-large
Voyage AI
1024$0.18$9.00$14.40
voyage-3
Voyage AI
1024$0.06$3.00$4.80
voyage-3-lite
Voyage AI
512$0.02$1.00$1.60
embed-english-v3
Cohere
1024$0.1$5.00$8.00
embed-multilingual-v3
Cohere
1024$0.1$5.00$8.00

About This Tool

The Embeddings Cost Calculator estimates indexing cost across OpenAI, Voyage AI, and Cohere embedding APIs. Enter average tokens per document, total document count, and the percentage of your corpus that updates each month. The tool computes initial indexing cost plus year-one update cost for each model.

Embedding pricing (April 2026)

OpenAI text-embedding-3-small: $0.02 per million tokens, 1536 dimensions, the cost-performance leader for general RAG. Text-embedding-3-large: $0.13/M, 3072 dims, marginally better recall. Voyage 3-large: $0.18/M, 1024 dims, the MTEB benchmark king for retrieval. Voyage 3-lite: $0.02/M, 512 dims, competitive with OpenAI's small at half the dimensions. Cohere Embed v3 (English and multilingual): $0.10/M, 1024 dims.

The hidden cost: re-embedding

Most teams budget for the initial index and forget update churn. If your corpus has 100K documents at 500 tokens each, the initial cost on text-embedding-3-small is $1. But if 10% of documents update monthly, that's another $1.20 per year on top. At enterprise scale (10M+ documents), the math reverses — update cost dominates the budget.

How to choose

Start with text-embedding-3-small. It's cheap, fast, and good enough for most retrieval tasks. Upgrade to large or Voyage 3-large only if your evals show meaningful recall improvement on your specific data. For multilingual corpora, Cohere's multilingual model is a strong default. For code retrieval, Voyage's code-tuned variants outperform general-purpose models.

Beyond embedding cost

Embedding is just the first line. You also need a vector database (storage + query cost), reranker (often Voyage Rerank or Cohere Rerank at $1-2/M tokens), and the LLM that consumes retrieved context. Use the vector database cost calculator for storage estimates. Compare against fine-tuning at RAG vs fine-tune calculator. For total stack budgeting, see AI monthly budget calculator. Estimate text size with token counter.

Frequently Asked Questions

What's the cheapest embedding model?
OpenAI's text-embedding-3-small at $0.02 per million tokens, tied with Voyage 3-lite. For most RAG use cases, 3-small is the default — strong recall at near-zero cost. A million-document corpus at 500 tokens each costs $10 to embed.
When should I pick text-embedding-3-large?
When recall is critical and the cost delta is acceptable. Large ($0.13/M) is 6.5x more expensive than small but scores ~5-10% higher on MTEB benchmarks. Worth it for high-value retrieval (legal, medical) or when you need 3072-dim vectors.
How is Voyage different from OpenAI embeddings?
Voyage 3-large tops the MTEB leaderboard for retrieval. It's pricier ($0.18/M) but produces noticeably better top-K results on technical and code corpora. Voyage 3 ($0.06/M) is a strong middle option. Worth A/B testing on your actual data.
Do embedding costs include re-embedding for updates?
Yes — every time a document changes you pay to re-embed it. Plan budget for re-indexing churn. If 5% of your 1M-document corpus updates monthly, that's 50K re-embeddings on top of the base index. Factor this into TCO.
How do dimensions affect storage and query cost?
Higher dimensions mean larger vectors and bigger index storage. 3072-dim vectors take 2x the memory of 1536-dim. Query latency also rises slightly. Pinecone, Qdrant, and Weaviate all charge by dimension × record count for storage.