Best embedding models, priced and ranked

What BEIR measures

BEIR (Benchmarking IR) is a zero-shot information-retrieval benchmark introduced by Thakur, Reimers, Rücklé, Srivastava, and Gurevych in 2021, spanning 15 English datasets across domains like biomedical, financial, scientific, and general-web search (MS MARCO, NQ, HotpotQA, FiQA, SciFact, and more). A model is given a query and must retrieve the most relevant passages from each dataset's corpus purely from its embeddings — no fine-tuning on the target dataset — which is what makes it "zero-shot": it measures how well an embedding generalizes to search tasks it wasn't trained for, which is exactly what happens when you plug an embedding model into a RAG pipeline over your own documents.

How to read this page

Unlike the chat-model benchmark pages on this site, the price axis here is real, direct pricing — not a proxy — since embedding APIs bill per input token with no separate output cost. Scores come from the official MTEB leaderboard's backend, which aggregates each model's results across BEIR's 15 tasks. A handful of newer paid APIs (marked with an asterisk) only have scores across part of BEIR's task set — we average what's available rather than hiding the model entirely, but treat those as a rougher signal than a full score. An absent score usually means the model hasn't been independently benchmarked yet, not that it's weak — commercial embedding APIs cost real money to run against a 15-dataset benchmark, so community coverage lags open-weight models that researchers can run for free on their own hardware.

Limitations

BEIR is English-only and retrieval-focused — a high score tells you a model is good at finding relevant documents for a query, not necessarily at classification, clustering, or multilingual tasks, which MTEB tracks separately. Embedding dimension isn't directly reflected in price here: a higher-dimensional embedding costs more to store and search (vector database size, index memory) even when the per-token API price is identical, so factor that in separately if storage cost matters for your use case. And as with every other data source on this site, prices and scores are joined automatically by matching model names across two independently-maintained projects — treat this as a strong starting point, not a substitute for checking the provider's own docs.

Frequently asked questions