OpenAI Cuts Embedding Costs 5x for RAG: $0.02 Instead of $0.10 per Million Tokens
Embeddings turn text into sets of numbers that RAG and search systems use to find semantically similar documents. OpenAI has released text-embedding-3-small, which costs $0.02 per million tokens versus $0.10 for its predecessor, ada-002, making it five times cheaper. On the MIRACL multilingual retrieval benchmark, it finds the right document in 44% of cases instead of 31.4%. If your RAG system runs on ada-002, the switch will pay off quickly, but you will have to re-embed your archive from scratch.
AI-processed from X: @OpenAI; edited by Hamidun News
If you are building document search, a chat over a knowledge base, or a RAG system, embeddings are the least visible and yet the most widespread line item in your costs. Each piece of text is first converted into a set of numbers, and then the system compares those sets to find the closest match in meaning.
On January 25, 2024, OpenAI announced an update to its developer platform and released a new model, text-embedding-3-small. A million tokens (roughly 750,000 English words) costs $0.02 to process, compared with $0.10 for the previous model, text-embedding-ada-002. That is five times cheaper, and quality did not drop; it went up.
The same announcement updated GPT-4 Turbo and cut the price of GPT-3.5 Turbo, but embeddings change the economics of search the most, so that is what we will focus on.
What Was Released
OpenAI now has two new models: text-embedding-3-small and text-embedding-3-large. The previous ada-002 appeared in December 2022 and remained the default choice for more than a year.
The smaller one, small, costs $0.02 per million tokens. To put that in perspective: if a database holds 100,000 documents of 1,000 tokens each, or 100 million tokens, one-time indexing costs $2 instead of $10.
Quality improved. On the English-language MTEB benchmark, small scores 62.3% against 61.0% for ada-002. The difference is more pronounced on the multilingual MIRACL benchmark: 44.0% versus 31.4%. Put simply, on non-English queries the system now finds the right document considerably more often.
The larger model, large, produces vectors of 3,072 numbers. It scores 64.6% on MTEB and 54.9% on MIRACL. It costs $0.13 per million tokens, slightly more than ada-002, but it is the most accurate option in the lineup.
How Vector Compression Works
The most practical novelty is the dimensions parameter. The longer the vector, the more memory it occupies in a vector database and the slower the search. Previously you could not choose the length: ada-002 always returned 1,536 numbers.
Now the vector can be shortened directly in the API request. According to OpenAI, large, cut down to 256 numbers, still outperforms the full-size ada-002 with its 1,536 numbers on MTEB. Think of a card catalog: instead of a cabinet of detailed cards, you keep short ones and still find what you need as well as before. For a database of millions of documents, that means a smaller storage bill and faster search.
What to Do in Practice
First: if you run production on ada-002 and handle a high volume of requests, small is the obvious candidate for replacement. Your embedding bill will drop by roughly a factor of five.
Second: migration is not just a matter of changing the model name in your code. Vectors from different models are incompatible, so you cannot search new vectors against an old index. You will have to recompute embeddings for all documents. This is a one-time operation, and at small's price it will be inexpensive, but it should be planned for and first run on a copy of the database.
Third: if retrieval over Russian or other non-English texts matters to you, look at MIRACL, not only MTEB. OpenAI's figures come from public tests, and results on your own data may differ. Before migrating, assemble your own set of typical queries and compare how often the old and new models retrieve the correct document.
Fourth: if you need maximum quality and price is secondary, choose large and experiment with dimensions right away. Find the vector length at which quality on your data still holds while the database stays compact.
There are limitations too. small's gain on MTEB is 1.3 points, which is modest, so the main reasons to migrate are cost and multilingual performance. The announcement does not show how the models behave on highly specialized texts such as medical or legal documents. The GPT-4 Turbo update and the GPT-3.5 Turbo price cut are also worth factoring into budget planning, but the decision about search should be made separately and based on your own measurements.
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026A swimming team captain handed ChatGPT three scheduled tasks and cut her routine work by a multiple
- Nov 21, 2026OpenAI GPT-6 Cyber Found About 35 Flaws in iPhone, Mac and Other Apple Devices
- Nov 21, 202622 Claude Agents Spent Six Weeks on Black Hole Problems — and Solved Two
Key Questions Answered
What Was Released?
OpenAI now has two new models: text-embedding-3-small and text-embedding-3-large. The previous ada-002 appeared in December 2022 and remained the default choice for more than a year. The smaller one, small, costs $0.02 per million tokens. To put that in perspective: if a database holds 100,000 documents of 1,000 tokens each, or 100 million tokens, one-time indexing costs $2 instead of $10. Quality improved. On the English-language MTEB benchmark, small scores 62.3% against 61.0% for ada-002.
How Vector Compression Works?
The most practical novelty is the dimensions parameter. The longer the vector, the more memory it occupies in a vector database and the slower the search. Previously you could not choose the length: ada-002 always returned 1,536 numbers. Now the vector can be shortened directly in the API request. According to OpenAI, large, cut down to 256 numbers, still outperforms the full-size ada-002 with its 1,536 numbers on MTEB.
What to Do in Practice?
First: if you run production on ada-002 and handle a high volume of requests, small is the obvious candidate for replacement. Your embedding bill will drop by roughly a factor of five. Second: migration is not just a matter of changing the model name in your code. Vectors from different models are incompatible, so you cannot search new vectors against an old index. You will have to recompute embeddings for all documents.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.