OpenAI Slashes Embedding Costs 99.8% and Unifies Five Models Into One

There’s a type of price cut that makes you do a double-take, and then there’s what OpenAI just dropped: 99.8% cheaper. The company’s new embedding model, text-embedding-ada-002, doesn’t just undercut its predecessors—it effectively renders the old pricing structure absurd while simultaneously outperforming the previous top dog, Davinci, across most benchmarks.

If you’ve been juggling separate models for text search, code search, and similarity, that headache disappears today. OpenAI has collapsed five distinct endpoints into this single model. One API call replaces a fragmented workflow that required developers to know which specialized model to use for which task. The unification alone would be notable, but pairing it with a fourfold increase in context length (now 8,192 tokens) and an embedding size slashed to 1,536 dimensions—one-eighth the size of Davinci-001—makes this a legitimate infrastructure upgrade rather than a mere refresh.

Notion is already on board, planning to move beyond basic keyword matching for its search feature. Kalendar AI offers a concrete performance metric worth paying attention to: using embeddings to match sales pitches against a dataset of 340 million profiles eliminated 40–56% of unwanted targeting compared to their old approach. That’s not a marginal improvement—it’s the difference between annoying prospects and actually reaching the right ones.

There is one asterisk. The new model doesn’t beat text-similarity-davinci-001 on the SentEval linear probing classification benchmark. For applications that train a lightweight linear layer on top of embeddings for classification, OpenAI suggests testing both and picking whichever performs better. Fair enough. But for the overwhelming majority of use cases—search, similarity, recommendations—the message is clear: you’d be burning money to stick with the old models.