Curiosity · AI Model
text-embedding-ada-002 (legacy)
text-embedding-ada-002 is OpenAI's second-generation text-embedding model, released December 2022 as a unified replacement for the earlier GPT-3 embedding endpoints. It produced 1536-dimensional dense embeddings with an 8k-token input limit, and for two years was the de facto default for retrieval-augmented generation, vector databases, and semantic search. Superseded in January 2024 by text-embedding-3-small and -3-large, which are both cheaper and stronger, but ada-002 remains widely deployed in legacy systems.
Model specs
- Vendor
- OpenAI
- Family
- OpenAI embeddings
- Released
- 2022-12
- Context window
- 8,191 tokens
- Modalities
- text
- Input price
- $0.1/M tok
- Output price
- n/a
- Pricing as of
- 2026-04-20
Strengths
- Stable API with years of production history
- Broad language coverage for its era
- Well-supported in every vector database
- Cheap — USD 0.10 per million tokens
Limitations
- Superseded by text-embedding-3-small (cheaper, stronger) and -3-large
- MTEB score ~10 points below 2024 frontier embedders
- 1536-dim fixed — no Matryoshka-style dimension shrinking
- English-biased compared with modern multilingual embedders
Use cases
- Maintaining existing RAG / vector DB pipelines built on ada-002
- Baseline reference for embedding benchmarks
- Gradual migration target → text-embedding-3-small
- Educational comparison with newer embedders
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MTEB English (avg, at release) | ≈61 | 2022-12 |
| BEIR retrieval (avg) | ≈49 | 2022-12 |
Frequently asked questions
What is text-embedding-ada-002?
OpenAI's December 2022 unified text-embedding model, producing 1536-dim dense embeddings. It was the default choice for RAG and semantic search through 2023.
Should I still use ada-002?
Only for legacy systems. New builds should use text-embedding-3-small or -3-large, which are cheaper and outperform ada-002 on MTEB and BEIR.
Is ada-002 deprecated?
OpenAI has flagged it as legacy but still serves the endpoint. Expect eventual deprecation — migrate when feasible.
Sources
- OpenAI — New embedding models — accessed 2026-04-20
- OpenAI — Embeddings overview — accessed 2026-04-20