Embedding benchmark refreshed: 25 models, 4 domains
Updated Embedding Models Benchmarks with nDCG@10 results from 200 questions across finance, legal, technical support and healthcare.
Embedding Benchmarks
Compare embedding models on 200 questions across finance, legal, technical support and healthcare. Explore AIMultiple's retrieval results, API prices and model details. See the methodology.
Benchmark results through October 1, 2026. API prices checked October 15, 2026.
TOP MODEL
Voyage 4 Large
nDCG@10 0.851
Best value
PPLX Embed v1 0.6B
$0.1263 benchmark cost
Largest model
PPLX Embed v1 4B
4.02B parameters
Coverage
25 models
4 domains · 200 questions
Embedding Models Benchmarks (nDCG@10)
Average retrieval score across 200 questions in four domains. Higher is better.
Leaderboard
All 25 models ranked by average nDCG@10 across finance, legal, technical support and healthcare.
Filter & Sort
# | Model | nDCG@10 | Finance | CUAD (legal) | TechQA (support) | MedRAG (healthcare) |
|---|---|---|---|---|---|---|
| 1 |  Voyage 4 Large Voyage AI | 0.851 | 0.838 | 0.704 | 0.887 | 0.978 |
| 2 |  Cohere Embed 5 Pro Cohere | 0.850 | 0.791 | 0.759 | 0.894 | 0.953 |
| 3 | Cohere Embed 5 Fast Cohere | 0.839 | 0.808 | 0.762 | 0.846 | 0.940 |
| 4 |  Gemini Embedding 2 Google | 0.830 | 0.824 | 0.731 | 0.796 | 0.967 |
| 5 | Voyage 4 Voyage AI | 0.829 | 0.782 | 0.700 | 0.855 | 0.980 |
| 6 | Voyage 4 Lite Voyage AI | 0.793 | 0.725 | 0.668 | 0.831 | 0.949 |
| 7 | Gemini Embedding 001 Google | 0.785 | 0.755 | 0.634 | 0.787 | 0.963 |
| 8 | Cohere Embed v4.0 Cohere | 0.772 | 0.723 | 0.626 | 0.785 | 0.953 |
| 9 | Voyage Code 4 Voyage AI | 0.770 | 0.824 | 0.687 | 0.616 | 0.952 |
| 10 |  Codestral Embed Mistral AI | 0.766 | 0.790 | 0.588 | 0.753 | 0.933 |
Page 1 of 3 | ||||||
Price ($/1M)$0.1200
Params-
ReleasedJan 2026
Finance
0.838
CUAD (legal)
0.704
TechQA (support)
0.887
MedRAG (healthcare)
0.978
Price ($/1M)-
Params-
ReleasedSep 2026
Finance
0.791
CUAD (legal)
0.759
TechQA (support)
0.894
MedRAG (healthcare)
0.953
Price ($/1M)-
Params-
ReleasedSep 2026
Finance
0.808
CUAD (legal)
0.762
TechQA (support)
0.846
MedRAG (healthcare)
0.940
Price ($/1M)$0.2000
Params-
ReleasedApr 2026
Finance
0.824
CUAD (legal)
0.731
TechQA (support)
0.796
MedRAG (healthcare)
0.967
Price ($/1M)$0.0600
Params-
ReleasedJan 2026
Finance
0.782
CUAD (legal)
0.700
TechQA (support)
0.855
MedRAG (healthcare)
0.980
Price ($/1M)$0.0200
Params-
ReleasedJan 2026
Finance
0.725
CUAD (legal)
0.668
TechQA (support)
0.831
MedRAG (healthcare)
0.949
Price ($/1M)$0.1500
Params-
ReleasedJul 2025
Finance
0.755
CUAD (legal)
0.634
TechQA (support)
0.787
MedRAG (healthcare)
0.963
Price ($/1M)-
Params-
ReleasedApr 2025
Finance
0.723
CUAD (legal)
0.626
TechQA (support)
0.785
MedRAG (healthcare)
0.953
Price ($/1M)$0.1200
Params-
ReleasedAug 2026
Finance
0.824
CUAD (legal)
0.687
TechQA (support)
0.616
MedRAG (healthcare)
0.952
Price ($/1M)$0.1500
Params-
ReleasedMay 2025
Finance
0.790
CUAD (legal)
0.588
TechQA (support)
0.753
MedRAG (healthcare)
0.933
Page 1 of 3
Embedding models benchmarks by release date
Each publisher's best average nDCG@10 by model release date. A step is a release that beat the publisher's earlier best; models without a release date are omitted. The chart shows 16 of 25 models.
Benchmark Cost vs Retrieval Accuracy
Total cost of embedding all 341,040 passages and 200 queries for each model, against average nDCG@10. This is the cost of running the benchmark, not a price per million tokens. Estimated cost: Cohere Embed v4.0. The nDCG@10 axis starts at 0.50.
Methodology
How Embedding Models Benchmarks are Built
We benchmarked embedding models on a fixed set of 200 questions: 50 each in finance, legal, technical support and healthcare. Each query searches its domain's full collection of text passages. We rank models by mean nDCG@10, which rewards placing recorded relevant passages near the top of the first ten results.
The four collections contain 341,040 text passages: chunks extracted from source documents, not a count of sentences or original documents. The questions were selected using BM25, and relevance labels are incomplete. These are development-set results, not a blind holdout evaluation. Embeddings were generated through hosted APIs; the cost chart shows what embedding the full benchmark cost each model, and API list prices appear separately in the leaderboard.
Recent Updates
Recent Updates
Latest changes to the Embedding Models Benchmarks, model coverage and AIMultiple benchmark methodology.
Update
Embedding benchmark refreshed: 25 models, 4 domains
Updated Embedding Models Benchmarks with nDCG@10 results from 200 questions across finance, legal, technical support and healthcare.
