Compare OpenAI text-embedding-3-small, Cohere, Voyage AI, and Jina pricing. Enter token volume → see exact monthly cost. Free, no signup.
Select model · Enter token volume · Results update live
Cost per million tokens, June 2026.
| Model | Price / 1M tokens | Dimensions | Best for |
|---|---|---|---|
| Jina Embeddings v2CHEAPEST | $0.018 | 768 | High-volume, cost-sensitive |
| OpenAI text-3-small | $0.020 | 1536 | Best quality/cost for most RAG |
| Cohere embed-v3 | $0.100 | 1024 | Multilingual, enterprise |
| OpenAI ada-002 | $0.100 | 1536 | Legacy, being phased out |
| Voyage AI voyage-2 | $0.120 | 1024 | Code search, technical docs |
| OpenAI text-3-large | $0.130 | 3072 | Max quality, high-precision RAG |
Calculate your embedding API cost for RAG systems, semantic search, and document indexing.
A SaaS company building a RAG system over 10,000 support articles, each split into 5 chunks of 300 tokens = 50,000 total chunks × 300 tokens = 15M tokens. OpenAI text-3-small: 15M × $0.020/1M = $0.30 for initial indexing. Monthly re-indexing (1,000 new articles/month = 5,000 chunks = 1.5M tokens): $0.03/month. Total annual embedding cost: $0.30 + $0.36 = $0.66/year. Embedding API costs are often the smallest line item in a RAG budget.
How we source embedding API pricing.
Sources: OpenAI embedding prices from OpenAI's API pricing page. Jina Embeddings pricing from Jina AI's pricing documentation. Cohere pricing from Cohere's API pricing page. Voyage AI pricing from their pricing page. All are pay-as-you-go, per-million-token rates with no minimum commitment.
Token counting methodology: Embedding APIs count tokens the same way as LLM APIs — approximately 1 token per 4 characters or 0.75 words in English. Non-English text and code may use more tokens per word. The calculator uses your entered "average tokens per document" directly — we recommend running a sample of your actual documents through the provider's tokenizer to get accurate estimates.
What is excluded: Vector database storage costs for the resulting embeddings (see Vector Database Cost Calculator), the cost of re-embedding documents when you switch providers, and query embedding costs (typically negligible — query embeddings are single-document calls, usually under 100 tokens, at the same per-token rate).
Self-hosting note: Sentence-transformers (open-source) models like all-MiniLM-L6-v2 (384 dims) can be self-hosted for near-zero API cost. Quality is lower than OpenAI text-3-small but acceptable for many use cases. On a $6/month Hetzner VPS, embedding throughput is approximately 1,000–2,000 documents per minute — adequate for most batch indexing workloads.
Jina Embeddings v2 at $0.018/M tokens is the cheapest commercial embedding API. OpenAI text-embedding-3-small at $0.020/M is a close second with superior quality and ecosystem support. For self-hosted options, sentence-transformers on a $6/month VPS provides near-zero API cost — the trade-off is slightly lower retrieval quality and operational maintenance overhead.
1536 dimensions (OpenAI text-3-small default) is the right choice for most RAG applications — it balances quality with storage cost. OpenAI's Matryoshka Representation Learning means you can truncate text-3-small vectors to 512 or 256 dimensions with less than 5% retrieval quality loss, cutting storage cost by 66–83%. Only use 3072-dimensional embeddings (text-3-large) for extremely precision-sensitive retrieval over large, noisy document sets.
Common chunk sizes: 100-word paragraph ≈ 133 tokens, 300-word article section ≈ 400 tokens, 1-page document ≈ 700 tokens, 10-page report ≈ 7,000 tokens. Most embedding models have an 8,191-token input limit — chunks longer than this must be split. Recommended chunk size: 200–500 tokens, with 50-token overlap between consecutive chunks for better context continuity in retrieval.
OpenAI text-3-small is almost always the right choice unless you're embedding millions of documents per day. The quality advantage over free models is meaningful (10–20% better retrieval accuracy on typical benchmarks), and at $0.020/1M tokens, the cost for most projects is under $1/month. Self-hosting makes sense at very high volumes (100M+ tokens/month = $2,000/month with OpenAI vs $20–50/month on a GPU server) or when your documents contain sensitive data that can't leave your infrastructure.
Matryoshka embeddings (used by OpenAI text-3-small and text-3-large) are trained so that the first N dimensions of a vector contain the most important semantic information. This means you can truncate a 1536-dim vector to 512 dims and retain approximately 95% of retrieval quality. The benefit: each stored vector takes 66% less space in your vector database, and HNSW index sizes shrink proportionally, reducing both storage cost and query latency. You don't pay less for the API call (still billed per input token) — savings are in vector DB storage and query speed.
Yes — embeddings from different models are not compatible. If you switch from OpenAI ada-002 to text-3-small, you must re-embed your entire document corpus and rebuild your vector index. For large datasets (millions of documents), this can take hours and incur significant API costs. This switching cost is a key reason to choose your embedding model carefully upfront. OpenAI text-3-small is a safe default — it's significantly better than ada-002 and unlikely to become obsolete soon.