text-embedding-3-small
OpenAIEmbeddingProprietaryOpenAI's efficient and affordable embedding model for converting text into 1536-dimensional vectors. Ideal for semantic search, clustering, and RAG with a 5x price reduction over ada-002.
Abilities
Use Cases
Available in Tools
Availability
How to Use
Pros
- Extremely affordable at $0.02/1M tokens — 5x cheaper than ada-002 and 6.5x cheaper than text-embedding-3-large
- Better performance than ada-002 despite being much cheaper — 62.3% vs 61.0% MTEB average
- Configurable dimensions (512–1536) — trade quality for storage and speed
- Simple API — just send text, receive vectors, no ML expertise needed
- Up to 8191 tokens input — handles reasonably long text chunks
Cons
- Lower quality than text-embedding-3-large (62.3% vs 64.6% MTEB) — not ideal when maximum retrieval accuracy matters
- Closed source — cannot self-host, all data must be sent to OpenAI
- Cannot fine-tune — stuck with generic embeddings
- Open-source alternatives (BGE, GTE) approaching comparable quality for free
What to Use It For
Perfect For
At $0.02/1M tokens it is the cheapest high-quality embedding option from OpenAI — ideal for large document collections
Good retrieval quality at a fraction of the cost of text-embedding-3-large
Good For
Configurable dimensions (512–1536) allow optimizing for speed in clustering workflows
Good semantic similarity at low cost makes it practical for deduplication of large collections
Not Recommended
2.3 percentage points lower MTEB score than text-embedding-3-large — use the large model when accuracy is critical
Try instead: text-embedding-3-large
All data must be sent to OpenAI servers — no offline or on-premise option
Try instead: Jina Embeddings v3
Do Not Use For
Embedding model only — converts text to vectors, cannot generate any text or responses
Try instead: GPT-4o
Cannot be fine-tuned — stuck with generic embeddings that may miss domain-specific nuances
Try instead: Jina Embeddings v3