Question

Why did the OpenAI embeddings endpoint change output dimensions

Answer 1

The OpenAI embeddings endpoint does not randomly “break” dimensions , the behavior changed because new embedding models introduced configurable output sizes and different default vector lengths.

Before in the old model
Text-embedded-ada-020 always would return 1536 dimensions

In the new model which  was introduced in 2024

Text-embedded-3-small =default 1536 dimensions

Text-embedded-3-large = default 3072 dimensions 

OpenAI added the dimension parameter
Now you can explicitly shorten the embeddings 

Json

{

  "model": "text-embedding-3-large",

  "input": "hello world",

  "dimensions": 1024

}

This means now the same output can return 1536,3072,1024,512 dimensions etc
I hope this helps .