Which embedding model is best for multilingual content?

Cohere embed-v4 is the strongest choice for multilingual content, especially non-Latin scripts (Arabic, Hindi, Japanese, Chinese). It supports 100+ languages and shows 15-20% quality improvement over OpenAI for these languages. Voyage domain-specific models are stronger for English domain-specific content.

Should I migrate from text-embedding-ada-002 to text-embedding-3-small?

Yes, strongly. ada-002 costs $0.10/M tokens vs $0.02/M for text-embedding-3-small - 5x cheaper, while also having better MTEB performance (62.3 vs 60.5). The re-embedding cost is small: 1 billion tokens costs $20 to re-embed. At equivalent ongoing volume, you pay back the migration cost in about 2-3 months.

Embedding Model Comparison: OpenAI vs Cohere vs Voyage vs Google vs Self-Hosted (June 2026)

Q: Which embedding model is cheapest?

Self-hosted BGE-M3 at ~$0.001/M tokens is cheapest overall but requires infrastructure. Among commercial APIs, OpenAI text-embedding-3-small batch at $0.01/M is cheapest, followed by OpenAI small standard and Voyage-3-lite both at $0.02/M.

Q: Which embedding model has the best accuracy?

Voyage's voyage-4-large (January 2026, mixture-of-experts) is the current accuracy leader: on Voyage's RTEB benchmark it beats OpenAI text-embedding-3-large by 14% and Cohere embed-v4 by 8.2% on NDCG@10. On the older MTEB Retrieval benchmark, the previous-generation voyage-3-large scored 68.9 and gemini-embedding-2-preview about 68.0, with OpenAI text-embedding-3-large at 64.6 and text-embedding-3-small at 62.3.

Q: Do Cohere and Voyage have free tiers?

Cohere has a rate-limited free tier (100 API calls/minute) with no token cap - suitable for development only. Voyage AI offers 200 million free tokens on the voyage-4 generation (plus voyage-context-3 and voyage-code-3) - more useful for prototyping; its older voyage-3.x models no longer include the free allocation. OpenAI has no free embedding tier after initial trial credits expire.

Q: Which model should I use for code search?

Voyage-code-3 is purpose-trained for code search, documentation retrieval, and GitHub issue search at $0.18/M tokens. For code at OpenAI, text-embedding-3-small works reasonably well but is not specialized. The quality difference is noticeable for retrieval within large codebases.

Q: What embedding model is best for a small RAG application?

For a small knowledge base under 100k documents with under 1k queries per day, OpenAI text-embedding-3-small with pgvector on an existing Postgres database keeps costs under $5 per month. The one-time embedding cost is $1 or less. See our RAG scenarios page for a detailed worked example.

Q: Is ada-002 vs text-embedding-3-small really that different?

text-embedding-3-small outperforms ada-002 on MTEB Retrieval (62.3 vs 60.5) while costing 5x less. For most RAG applications, retrieval quality improves slightly and cost drops dramatically. The only reason to stay on ada-002 is if you have an existing production index and want to defer the re-embedding cost.

Side-by-side comparison of every major embedding model. Price, dimensions, MTEB score, context window, batch discount, free tier, and best-fit use case.

Verified June 2026

Full Model Comparison Table

Model	Provider	$/M std	$/M batch	Dims	Context	MTEB	Free tier	Best for
text-embedding-3-small	OpenAI	$0.020	$0.010	1,536	8,191	62.3	-	General-purpose RAG, low cost
text-embedding-3-large	OpenAI	$0.130	$0.065	3,072	8,191	64.6	-	High-accuracy retrieval
text-embedding-ada-002	OpenAI	$0.100	No batch	1,536	8,191	60.5	-	Legacy - migrate to 3-small
embed-v4	Cohere	$0.120	No batch	1,536	128,000	55	100 calls/min	Multilingual content (100+ languages)
voyage-4-large	Voyage AI	$0.120	$0.080	1,024	32,000	-	200M tokens	State-of-the-art retrieval (MoE)
voyage-4	Voyage AI	$0.060	$0.040	1,024	32,000	-	200M tokens	Best price-to-accuracy ratio
voyage-4-lite	Voyage AI	$0.020	$0.013	1,024	32,000	-	200M tokens	High-volume, cost-sensitive
gemini-embedding-2-preview	Google	$0.200	No batch	3,072	8,192	68	-	Multimodal, Google ecosystem
gemini-embedding-001	Google	$0.150	No batch	3,072	2,048	65.4	-	Stable GA, Google ecosystem
Titan Text Embeddings V2	Amazon Bedrock	$0.020	No batch	1,024	8,192	62.8	-	AWS-native apps, compliance
BGE-M3 (self-hosted)	Self-Hosted	$0.001	$0.001	1,024	8,192	66.5	-	High-volume, privacy-sensitive

Green = best in category, amber = watch out, purple = top MTEB. MTEB Retrieval average where publicly available. Context window in tokens. Verified June 2026.

By Scenario: Which Model Should You Pick?

Small RAG bot, under 100k documents, budget-first

OpenAI text-embedding-3-small + pgvector

Under $5/month total. One-time embedding cost for 50M tokens is $1.00. pgvector is near-zero marginal cost on existing Postgres.

Provider details

Best accuracy for production RAG

Voyage voyage-4

Approaches voyage-3-large quality at a mid-sized model's cost. At $0.06/M standard or $0.04/M batch, the accuracy premium is affordable for most production apps, with 200M free tokens to start.

Provider details

Multilingual content, 100+ languages

Cohere embed-v4

15-20% quality improvement for non-Latin scripts over OpenAI. Its 128K-token context handles long documents with minimal chunking.

Provider details

Maximum retrieval accuracy, cost secondary

Voyage voyage-4-large

Tops Voyage's RTEB benchmark (+14% vs OpenAI 3-large on NDCG@10) with a mixture-of-experts design. At $0.12/M, index with large and query with voyage-4-lite ($0.02/M) using the shared-vector-space feature.

Provider details

AWS-native application, compliance required

Amazon Titan V2 via Bedrock

Keeps data within AWS VPC. Binary embedding option for 32x storage reduction. Integrates with OpenSearch Serverless and SageMaker.

Provider details

Over 15 million tokens per month

Self-hosted BGE-M3

At A100 spot rates, $0.001/M tokens vs $0.02/M for OpenAI small. Breaks even at roughly 15M tokens/month. Requires DevOps investment.

Provider details

Code search or documentation RAG

Voyage voyage-code-3

Purpose-trained on code repositories at $0.18/M. Meaningfully better than general-purpose models for code retrieval, with 200M free tokens to evaluate it.

Provider details

Google ecosystem, GCP billing

Google gemini-embedding-001

Stable GA model, $0.15/M, strong MTEB at 65.4. MRL support for dimension reduction. Integrates naturally with Vertex AI, BigQuery, and Cloud Storage.

Provider details

Head-to-Head Comparisons

OpenAI small vs Cohere embed-v4

text-embedding-3-small

Price: $0.02/M

Dims: 1536

MTEB: 62.3

Context: 8,191

Free: Trial only

Cheaper; better for English

embed-v4

Price: $0.12/M

Dims: 1536

MTEB: 55.0

Context: 128K

Free: 100 calls/min

Better multilingual; 128K context

OpenAI large vs Voyage 4 large

text-embedding-3-large

Price: $0.13/M

Dims: 3072

MTEB: 64.6

Context: 8,191

Free: Trial only

Mature ecosystem

voyage-4-large

Price: $0.12/M

Dims: 1024

MTEB: RTEB-led

Context: 32,000

Free: 200M lifetime

MoE; +14% vs OAI 3-large on RTEB

Voyage 4 vs Gemini embedding-001

voyage-4

Price: $0.06/M

Dims: 1024

MTEB: n/a

Context: 32,000

Free: 200M lifetime

Best accuracy-to-cost ratio

gemini-embedding-001

Price: $0.15/M

Dims: 3072

MTEB: 65.4

Context: 2,048

Free: AI Studio free tier

GCP ecosystem; MRL dims

ada-002 vs text-embedding-3-small (migration)

ada-002 (legacy)

Price: $0.10/M

Dims: 1536

MTEB: 60.5

Context: 8,191

Free: N/A

No - expensive, worse quality

text-embedding-3-small

Price: $0.02/M

Dims: 1536

MTEB: 62.3

Context: 8,191

Free: Trial only

Yes - 5x cheaper, better quality

Frequently Asked Questions

Which embedding model is cheapest?

Self-hosted BGE-M3 at ~$0.001/M tokens is cheapest overall. Among commercial APIs, OpenAI text-embedding-3-small batch at $0.01/M is cheapest, followed by OpenAI small standard and voyage-4-lite both at $0.02/M.

Which embedding model has the best accuracy?

Voyage's voyage-4-large (January 2026, MoE) leads its RTEB benchmark, beating OpenAI text-embedding-3-large by 14% and Cohere embed-v4 by 8.2% on NDCG@10. On the older MTEB Retrieval benchmark, voyage-3-large scored 68.9, gemini-embedding-2-preview ~68.0, and OpenAI text-embedding-3-large 64.6.

Do Cohere and Voyage have free tiers?

Cohere has a rate-limited free tier (100 API calls/minute). Voyage AI offers 200 million free tokens on the voyage-4 generation (plus voyage-context-3 and voyage-code-3); its older voyage-3.x models no longer include the free allocation. OpenAI has no free embedding tier after trial credits expire.

Which model is best for multilingual content?

Cohere embed-v4 is strongest for multilingual content, especially non-Latin scripts. It shows 15-20% quality improvement over OpenAI for Arabic, Hindi, Japanese, and Chinese.

Which model should I use for code search?

Voyage-code-3 is purpose-trained for code search at $0.18/M tokens. For code with OpenAI, text-embedding-3-small works but is not specialized.

Should I migrate from ada-002 to text-embedding-3-small?

Yes. ada-002 costs $0.10/M vs $0.02/M for text-embedding-3-small - 5x cheaper with better quality. Re-embedding 1B tokens costs $20 and pays back in 2-3 months at equivalent volume.

What embedding model is best for a small RAG application?

OpenAI text-embedding-3-small with pgvector on an existing Postgres database costs under $5/month for a 100k-document knowledge base with low query volume.

Is ada-002 vs text-embedding-3-small really that different?

text-embedding-3-small outperforms ada-002 (MTEB 62.3 vs 60.5) while costing 5x less. The only reason to stay on ada-002 is to defer a re-embedding pass on an existing index.

Full cost calculator

See Year 1 total including storage for every provider

Optimization techniques

Batch API, MRL dimensions, chunking, caching

Disclaimer: Independent resource. Not affiliated with any provider listed. Prices and MTEB scores verified June 2026. MTEB scores from the public MTEB leaderboard (Retrieval task average). Always verify pricing on each provider's own site before making decisions.