Curiosity · AI Model
OpenAI text-embedding-3-small
text-embedding-3-small is OpenAI's throughput-tuned embedding model — 1536 default dimensions and roughly five times cheaper than text-embedding-3-large while still beating the legacy ada-002. It is the right default when you are embedding tens of millions of chunks or running cost-sensitive semantic search at scale.
Model specs
- Vendor
- OpenAI
- Family
- text-embedding-3
- Released
- 2024-01
- Context window
- 8,191 tokens
- Modalities
- text
- Input price
- $0.02/M tok
- Output price
- n/a
- Pricing as of
- 2026-04-20
Strengths
- ≈5× cheaper per million tokens than text-embedding-3-large
- Strong quality — still beats the legacy text-embedding-ada-002
- Matryoshka truncation lets you trade a few MTEB points for far smaller vectors
- Fast enough to embed millions of documents per hour
Limitations
- Lower MTEB score than text-embedding-3-large — use the large model for high-stakes retrieval
- Closed weights — cannot self-host
- Multilingual quality is decent but lags Cohere embed-v3 and Voyage-3 on some locales
Use cases
- Large-scale corpus ingestion for RAG
- Semantic search with tight cost ceilings
- Deduplication and near-duplicate detection
- Recommendation features in consumer apps
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MTEB (English, avg) | ≈62.3 | 2024-01 |
| MIRACL (multilingual) | ≈44.0 | 2024-01 |
Frequently asked questions
What is text-embedding-3-small?
text-embedding-3-small is OpenAI's cost-optimised embedding model, producing 1536-dimensional vectors. It replaces text-embedding-ada-002 for most workloads and is the recommended default for bulk ingestion and cost-sensitive RAG.
How does text-embedding-3-small compare with ada-002?
It is both higher-quality on MTEB and cheaper per million tokens than ada-002. OpenAI recommends new projects use text-embedding-3-small or text-embedding-3-large rather than the legacy ada-002 model.
How much does text-embedding-3-small cost?
As of April 2026, text-embedding-3-small costs roughly USD 0.02 per million input tokens on the OpenAI API, making it viable for embedding very large corpora.
Can I shrink text-embedding-3-small vectors?
Yes — the model supports a dimensions parameter via Matryoshka representation learning, letting you request shorter vectors (for example 512 or 256 dims) to save storage with modest quality impact.
Sources
- OpenAI — New embedding models — accessed 2026-04-20
- OpenAI — Embeddings guide — accessed 2026-04-20