Best embedding model APIs
Text-embedding models power search and RAG, and they're priced per input token only — no output side. This ranks 15 models on a single dated snapshot of the MTEB (multilingual mean task score) (a mean across retrieval, classification, clustering, STS and reranking tasks), then divides the score by each API's price per 1M tokens to find the best quality per dollar. Open-weight models you self-host are listed too — no per-token bill, so no value figure.
Best value: jina-embeddings-v3 (Jina AI) — 58.37 MTEB at $0.02/1M = 2,918.5 points per dollar.
Best value: MTEB points per dollar
API models with a sourced price, ranked by score ÷ price per 1M input tokens.
| # | Model | MTEB | Dims | Price /1M | MTEB / $ | |
|---|---|---|---|---|---|---|
| 1 | jina-embeddings-v3 · Jina AI | 58.37 | 1,024 | $0.02 | 2,918.5 | |
| 2 | voyage-3.5 · Voyage AI | 58.46 | 1,024 | $0.06 | 974.3 | |
| 3 | gemini-embedding-001 · Google | 68.37 | 3,072 | $0.15 | 455.8 | |
| 4 | text-embedding-3-large · OpenAI | 58.96 | 3,072 | $0.13 | 453.5 |
Value = MTEB score ÷ price per 1M input tokens. Prices are per-model and each links to its source; scores come from one dated leaderboard so the ranking is internally consistent. Cheaper isn't automatically better — a two-point MTEB gap can matter more than a price gap for hard retrieval. Dimensions drive your vector-store cost too.
All models by MTEB score
The full board, highest quality first. Open-weight rows have no API price — you run them yourself.
| # | Model | Type | MTEB | Dims | Params | License | Price /1M |
|---|---|---|---|---|---|---|---|
| 1 | KaLM-Embedding-Gemma3-12B · Tencent | open weights | 72.32 | 3,840 | 11.76B | Tencent Community License | self-host |
| 2 | Qwen3-Embedding-8B · Alibaba | open weights | 70.58 | 4,096 | 8B | Apache-2.0 | self-host |
| 3 | Seed1.6-embedding · ByteDance | API | 70.26 | 1,536 | — | Proprietary | self-host |
| 4 | llama-embed-nemotron-8b · NVIDIA | open weights | 69.46 | 4,096 | 8B | NVIDIA Open Model License | self-host |
| 5 | Qwen3-Embedding-4B · Alibaba | open weights | 69.45 | 2,560 | 4B | Apache-2.0 | self-host |
| 6 | gemini-embedding-001 · Google | API | 68.37 | 3,072 | — | Proprietary | $0.15 |
| 7 | Octen-Embedding-8B · Octen | open weights | 67.85 | 4,096 | 8B | — | self-host |
| 8 | Qwen3-Embedding-0.6B · Alibaba | open weights | 64.34 | 1,024 | 0.6B | Apache-2.0 | self-host |
| 9 | multilingual-e5-large-instruct · Microsoft | open weights | 63.22 | 1,024 | 560M | MIT | self-host |
| 10 | gte-Qwen2-7B-instruct · Alibaba | open weights | 62.51 | 3,584 | 7B | Apache-2.0 | self-host |
| 11 | text-multilingual-embedding-002 · Google | API | 62.16 | 768 | — | Proprietary | self-host |
| 12 | bge-m3 · BAAI | open weights | 59.56 | 1,024 | 568M | MIT | self-host |
| 13 | text-embedding-3-large · OpenAI | API | 58.96 | 3,072 | — | Proprietary | $0.13 |
| 14 | voyage-3.5 · Voyage AI | API | 58.46 | 1,024 | — | Proprietary | $0.06 |
| 15 | jina-embeddings-v3 · Jina AI | APIopen weights | 58.37 | 1,024 | 570M | CC BY-NC 4.0 | $0.02 |
6 of 15 are hosted APIs, 10 ship open weights.MTEB (multilingual mean task score) · scores verified 2026-05-17 ·leaderboard source. One dated snapshot, curated by hand — MTEB versions and task mixes drift between sources, so mixing them adds noise.
FAQ
What is the best value embedding model API?
jina-embeddings-v3 (Jina AI) leads on quality per dollar: 58.37 MTEB (multilingual mean task score) at $0.02 per 1M tokens = 2,918.5 MTEB points per dollar.
Why is embedding pricing different from chat model pricing?
Text-embedding models are billed per input token only — there is no output or generation side. Value here is the MTEB score divided by the price per 1M input tokens, and your vector-store cost also scales with each model's output dimensions.
Which embedding API is cheapest?
jina-embeddings-v3 is the cheapest sourced API at $0.02 per 1M input tokens. Cheaper isn't automatically better — a two-point MTEB gap can matter more than a price gap on hard retrieval.