What replaced text-embedding-ada-002 in OpenAI's lineup?

OpenAI's text-embedding-ada-002 was the default embedding model from late 2022 through early 2024. It produced 1536-dimensional vectors at a single fixed dimension and a single fixed price. In January 2024, OpenAI released two replacements: text-embedding-3-small and text-embedding-3-large. The small model also produces 1536-dimensional vectors but at half the price of ada-002. The large model produces 3072-dimensional vectors at a higher price than ada-002 but with substantially better retrieval quality. Both new models score 5 to 10 points higher than ada-002 on MTEB, the standard embeddings benchmark. Both also support Matryoshka truncation through the 'dimensions' API parameter, which lets you ask for a shorter vector (say, 512 or 1024 dimensions) for storage savings without re-embedding. The ada-002 is still available, but the new lineup is the recommended default for any new build.

text-embedding-ada-002 was the default OpenAI embedding model from late 2022. In January 2024 it was superseded by the text-embedding-3 family: text-embedding-3-small and text-embedding-3-large. The small model is 1536-dimensional, the large is 3072-dimensional, and both deliver a 5 to 10 point lift over ada-002 on the MTEB benchmark. The pricing changed in the user's favor. The text-embedding-3-small ships at $0.02 per million tokens, roughly half the ada-002 rate. The text-embedding-3-large is the premium option at $0.13 per million tokens, used when you need the quality lift and the storage cost is acceptable. The headline new capability is **Matryoshka truncation**. Both models are trained so that any prefix of the vector is itself a usable, lower-dimensional embedding. The API exposes this through a 'dimensions' parameter. You can ask for 512 or 1024 dimensions and get a meaningful vector at that size. The ada-002 had no such flexibility. You got 1536 or nothing. The ada-002 is still served but is no longer recommended for new builds.