Teams paying six figures a year for vector search infrastructure just had their cost assumptions cut in half overnight.

Most embedding pipelines are expensive and quietly inaccurate

Retrieval-augmented generation systems live or die by embedding quality, and most teams have been trading accuracy for cost without a clean alternative. Reindexing entire corpora every time a better model appeared was expensive enough that most teams simply didn’t.

The models got faster, cheaper, and more accurate at the same time

New embedding models and API updates gives developers two new text embedding models — text-embedding-3-small and text-embedding-3-large — plus a native dimensions parameter that lets you shrink vector size without retraining, and an updated API that now exposes usage tokens per embedding call. You paste in your text, set your target dimensions, and get back a smaller, cheaper vector that benchmarks higher on MTEB than the previous generation.

Search and RAG engineers are going to feel this first

• ML engineers running RAG pipelines who need to cut Pinecone or Weaviate storage costs without sacrificing retrieval precision.
• Backend developers billing API costs to clients who can now re-price contracts with a 50% reduction in embedding spend.
• Data teams maintaining semantic search over large document libraries who can finally re-embed without a budget approval cycle.

The dimensions parameter is the detail most people will miss on first read, and it is the one that matters most for teams constrained by vector database storage limits.

Cohere and Voyage have been winning cost-sensitive deals for six months

OpenAI’s previous embedding models were losing ground to Cohere’s embed-v3 and Voyage’s models on both price and MTEB scores, which pushed a measurable share of new RAG projects toward alternatives. With this release, OpenAI embedding models re-enter the top tier of the MTEB leaderboard while cutting per-token pricing, which means the switching calculus for teams mid-migration just changed.

What you can do with it

• Re-embed existing corpora at lower cost with no accuracy penalty.
• Reduce Pinecone or Qdrant storage by truncating dimensions natively.
• Track exact token consumption per call using the new usage field.
• Replace text-embedding-ada-002 with a one-line model swap.

Pricing is $0.00002 per 1K tokens for the small model and $0.00013 per 1K tokens for the large model.

The dimensions parameter only works natively with these two new models, so any pipeline still on ada-002 requires a model migration before it can benefit.

Cohere Embed v3 remains competitive on multilingual tasks. Voyage AI is still the default choice for code-heavy retrieval workloads where domain specificity outweighs general benchmark scores.

The embedding market just reset and most teams haven’t noticed yet

We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.