text-embedding-3-small

text-embedding-3-small launched on 25 January 2024 as the efficient half of OpenAI's second embedding generation. It produces 1,536-dimension vectors — the same length as the ada-002 model it supersedes — and, like its larger sibling, supports truncation through the dimensions parameter. On OpenAI's published figures it beats ada-002 on both benchmarks quoted: MIRACL multilingual retrieval rises from 31.4% to 44.0%, MTEB from 61.0% to 62.3%. The English gain is modest; the multilingual gain is not. What makes this model unusual in a catalogue of AI pricing is the direction the cost moved. At $0.02 per million tokens it is one fifth the price of ada-002, which still sits on the price list at $0.10 — a better model, at the same vector size, for a fifth of the money, and both are still sold today. Vendors normally retire the model that loses on both axes. This one is still there, and anyone who wired ada-002 into a pipeline in 2023 and never revisited the choice is paying five times more than they need to.

▸Videos

No videos yet.

▸Reviews