Text embeddings
Ranked by an independent measurement first, then by what it costs per 1M input tokens across sellers. Absence of a score is not a zero; absence of a price is not free.
2 undominated of 7 scored and priced
MTEB(eng, v2) only. Other task scores are shown separately.
The frontier
- Qwen3-Embedding-8B75.22 · $0.010 per 1M input tokens
- EmbeddingGemma 300M69.67 · $0.0020 per 1M input tokens
MTEB(eng, v2)
Sorted by the MTEB(eng, v2) mean over 41 tasks. A model with fewer than 41 tasks has no mean and is listed as partial with its count.
| Rank | Model | MTEB(eng, v2) mean | Price | Sellers | Spread | Verdict |
|---|---|---|---|---|---|---|
| — | QZhou-Embedding Kingsoft OW | 75.97 | no traceable price | — | — | unpriced |
| — | Qwen3-Embedding-8B Alibaba Qwen OW | 75.22 | $0.010 DeepInfra | 3 | 10× | undominated |
| — | Jasper-Token-Compression-600M infgrad OW | 74.75 | no traceable price | — | — | unpriced |
| — | Qwen3-Embedding-4B Alibaba Qwen OW | 74.60 | $0.020 DeepInfra | 2 | 1× | dominated by Qwen3-Embedding-8B |
| — | Seed1.6-embedding ByteDance | 74.07 | no traceable price | — | — | unpriced |
| — | gemini-embedding-001 Google | 73.30 | $0.150 OpenRouter | 1 | — | dominated by Qwen3-Embedding-8B |
| — | jina-embeddings-v5-text-nano Jina AI OW | 71.11 | no traceable price | — | — | unpriced |
| — | Qwen3-Embedding-0.6B Alibaba Qwen OW | 70.47 | $0.010 DeepInfra | 1 | — | dominated by Qwen3-Embedding-8B |
| — | EmbeddingGemma 300M Google OW | 69.67 | $0.0020 DeepInfra | 1 | — | undominated |
| — | text-embedding-005 Google | 69.60 | no traceable price | — | — | unpriced |
| — | text-multilingual-embedding-002 Google | 67.67 | no traceable price | — | — | unpriced |
| — | mxbai-embed-large-v1 Mixedbread OW | 66.26 | no traceable price | — | — | unpriced |
| — | multilingual-e5-large-instruct Microsoft (intfloat) OW | 65.56 | $0.010 DeepInfra | 1 | — | dominated by EmbeddingGemma 300M |
| — | nomic-embed-text-v2-moe Nomic OW | 63.64 | no traceable price | — | — | unpriced |
| — | snowflake-arctic-embed-l-v2.0 Snowflake OW | 63.60 | no traceable price | — | — | unpriced |
| — | granite-embedding-english-r2 IBM OW | 62.84 | no traceable price | — | — | unpriced |
| — | voyage-3 Voyage AI | 61.38 | no traceable price | — | — | unpriced |
| — | voyage-3-lite Voyage AI | 60.44 | no traceable price | — | — | unpriced |
| — | all-MiniLM-L6-v2 Sentence Transformers OW | 59.03 | $0.0050 DeepInfra | 2 | 1× | dominated by EmbeddingGemma 300M |
| — | paraphrase-multilingual-MiniLM-L12-v2 Sentence Transformers OW | 56.42 | no traceable price | — | — | unpriced |
| — | Nemotron-3-Embed-1B NVIDIA OW | partial, 0 of 41 | free OpenRouter | 2 | — | unrated |
| — | pplx-embed-v1-0.6b Perplexity | partial, 0 of 41 | $0.0040 OpenRouter | 1 | — | unrated |
| — | bge-base-en-v1.5 BAAI OW | partial, 33 of 41 | $0.0050 DeepInfra | 2 | 1× | unrated |
| — | e5-base-v2 Microsoft (intfloat) OW | partial, 33 of 41 | $0.0050 DeepInfra | 2 | 1× | unrated |
| — | all-MiniLM-L12-v2 Sentence Transformers OW | partial, 33 of 41 | $0.0050 DeepInfra | 2 | 1× | unrated |
| — | all-mpnet-base-v2 Sentence Transformers OW | partial, 33 of 41 | $0.0050 DeepInfra | 2 | 1× | unrated |
| — | multi-qa-mpnet-base-dot-v1 Sentence Transformers OW | partial, 0 of 41 | $0.0050 DeepInfra | 2 | 1× | unrated |
| — | gte-base Alibaba DAMO (thenlper) OW | partial, 21 of 41 | $0.0050 DeepInfra | 2 | 1× | unrated |
| — | bge-large-en-v1.5 BAAI OW | partial, 33 of 41 | $0.010 DeepInfra | 2 | 1× | unrated |
| — | bge-m3 BAAI OW | partial, 27 of 41 | $0.010 DeepInfra | 2 | 1× | unrated |
| — | e5-large-v2 Microsoft (intfloat) OW | partial, 33 of 41 | $0.010 DeepInfra | 2 | 1× | unrated |
| — | multilingual-e5-large Microsoft (intfloat) OW | partial, 23 of 41 | $0.010 DeepInfra | 2 | 1× | unrated |
| — | gte-large Alibaba DAMO (thenlper) OW | partial, 21 of 41 | $0.010 DeepInfra | 2 | 1× | unrated |
| — | text-embedding-3-small OpenAI | partial, 6 of 41 | $0.020 OpenRouter | 1 | — | unrated |
| — | pplx-embed-v1-4b Perplexity | partial, 0 of 41 | $0.030 OpenRouter | 1 | — | unrated |
| — | Nemotron-3-Embed-8B NVIDIA OW | partial, 0 of 41 | $0.035 DeepInfra | 1 | — | unrated |
| — | mistral-embed Mistral | partial, 0 of 41 | $0.100 OpenRouter | 1 | — | unrated |
| — | text-embedding-ada-002 OpenAI | partial, 0 of 41 | $0.100 OpenRouter | 1 | — | unrated |
| — | text-embedding-3-large OpenAI | partial, 10 of 41 | $0.130 OpenRouter | 1 | — | unrated |
| — | codestral-embed Mistral | partial, 0 of 41 | $0.150 OpenRouter | 1 | — | unrated |
| — | gemini-embedding-2 Google | partial, 0 of 41 | $0.200 OpenRouter | 1 | — | unrated |
| — | gte-modernbert-base Alibaba NLP OW | partial, 27 of 41 | no traceable price | — | — | no data |
| — | gte-Qwen2-7B-instruct Alibaba NLP OW | partial, 33 of 41 | no traceable price | — | — | no data |
| — | Titan Text Embeddings V2 Amazon | partial, 27 of 41 | no traceable price | — | — | no data |
| — | embed-english-v3.0 Cohere | partial, 21 of 41 | no traceable price | — | — | no data |
| — | embed-multilingual-v3.0 Cohere | partial, 21 of 41 | no traceable price | — | — | no data |
| — | embed-v4.0 Cohere | partial, 0 of 41 | no traceable price | — | — | no data |
| — | jina-embeddings-v3 Jina AI OW | partial, 30 of 41 | no traceable price | — | — | no data |
| — | jina-embeddings-v4 Jina AI OW | partial, 19 of 41 | no traceable price | — | — | no data |
| — | jina-embeddings-v5-text-small Jina AI OW | partial, 37 of 41 | no traceable price | — | — | no data |
| — | nomic-embed-text-v1.5 Nomic OW | partial, 33 of 41 | no traceable price | — | — | no data |
| — | voyage-3-large Voyage AI | partial, 3 of 41 | no traceable price | — | — | no data |
| — | voyage-3.5 Voyage AI | partial, 22 of 41 | no traceable price | — | — | no data |
| — | voyage-4 Voyage AI | partial, 0 of 41 | no traceable price | — | — | no data |
| — | voyage-4-large Voyage AI | partial, 0 of 41 | no traceable price | — | — | no data |
| — | voyage-4-lite Voyage AI | partial, 0 of 41 | no traceable price | — | — | no data |
| — | voyage-code-3 Voyage AI | partial, 29 of 41 | no traceable price | — | — | no data |
Worth shopping
The same model at more than one seller, dearest divided by cheapest, over distinct sellers at the reference unit. A vendor's own tiers are not competition.
| Model | Spread | Sellers | Cheapest |
|---|---|---|---|
| Qwen3-Embedding-8B | 10× | 3 | $0.010 DeepInfra |
How this page is built
- Quality. The mean over the 41 tasks of MTEB(eng, v2), computed here from the CC0 per-task result files. Scores reach that repository as pull requests, usually from the vendor.
- Price. Vendor pages read by hand, plus the public price feeds of OpenRouter and DeepInfra. Every figure carries the page it came from and the day it was read. Community aggregations are used to find a vendor page, never as the number.
- Normalisation. One reference unit, per 1M input tokens. Conversions are printed next to each figure on the model page. A price stated under conditions that cannot be converted is shown and excluded from the ranking rather than guessed at.
- Verdict. Computed within each model's selected benchmark task. A competitor must be at least as good and at least as cheap, with a strict advantage on one axis. The headline and plot cover MTEB(eng, v2) only; an alternate task score does not inherit a task-specific price.
MTEB results, https://github.com/embeddings-benchmark/results, released under CC0-1.0; MTEB(eng, v2) means computed from the per-task files as described in data/modalities/embeddings/benchmark.json.
Questions this page answers
What does the score on this page measure?
It is the mean over the 41 tasks of MTEB(eng, v2), computed from the CC0 results repository rather than copied from a leaderboard. MTEB publishes no confidence interval, so none is shown, and a model missing any of the 41 tasks gets no mean at all.
What is a price per 1M input tokens here?
Every seller states a price in its own unit. Each figure is normalised to one reference unit, per 1M input tokens, with the rule printed beside it on the model page, and nothing is rescaled across resolutions or quality tiers. A figure that cannot be converted without inventing something is shown, not ranked. So is a quote for one part of the bill, such as the output image when input tokens or a search tool are charged on top.
Why do so many models have no price?
13 of these models are scored but have no comparable price that traces to a page we have read. Missing is not free, and an unpriced model is never called cheap. Most vendors in this market publish prices as JavaScript pages or not at all.
What does undominated mean?
Within the same selected benchmark task, no other scored and priced model is at least as good and at least as cheap, with a strict advantage on one axis. Separate benchmark tasks never share a frontier. Missing scores or prices do not qualify a model.