text-embedding-3-large vs text-embedding-3-small

Direct comparison between openai and openai for embedding models. Specs, prices, strengths, weaknesses — decide on facts.

At a glance

  • · Cheaper: text-embedding-3-small
  • · text-embedding-3-large: 3072 dimensions — top accuracy on MTEB benchmarks
  • · text-embedding-3-small: One of the cheapest embedding models per million tokens

Spec comparison

Spectext-embedding-3-largetext-embedding-3-small
Provideropenaiopenai
Input per 1M tokens$0.13 / 1M$0.02 / 1M
Output per 1M tokens$0.00 / 1M$0.00 / 1M
Context window8,191 tok8,191 tok

text-embedding-3-large — strengths

  • +3072 dimensions — top accuracy on MTEB benchmarks
  • +Dimension can be shortened via API (Matryoshka) — save storage without re-embedding
  • +Multilingual, strong on cross-language retrieval

text-embedding-3-small — strengths

  • +One of the cheapest embedding models per million tokens
  • +Clearly better than the old ada-002 at lower price
  • +Default pick for RAG prototypes and mid-sized corpora

text-embedding-3-large — weaknesses

  • Pricier than 3-small, often with no noticeable quality edge on small corpora
  • Full 3072-dim is storage-intensive

text-embedding-3-small — weaknesses

  • Behind 3-large on the hardest MTEB tasks
Details: text-embedding-3-largeDetails: text-embedding-3-small