Cohere vs Voyage vs OpenAI Embeddings (2026): The Honest RAG Comparison

Embedding cost is usually a one-time-per-document expense (you embed once at ingest, then query against the vector store), so the headline $/1M number understates how cheap embeddings actually are at most operational scales. A 1M-document corpus averaging 500 tokens per doc = 500M tokens to embed once.

**OpenAI text-embedding-3-small** at $0.02/1M × 500M = **$10 one-time** for the full corpus. The Batch API at 50% off = **$5**. text-embedding-3-large at $0.13/1M × 500M = **$65** (or $32.50 batched). For most non-trivial RAG systems, the embedding bill is rounding error compared to the LLM inference bill that comes after.

**Voyage AI voyage-3-lite** at $0.02/1M × 500M = **$10 one-time**. voyage-3 at $0.06/1M = **$30**. voyage-3-large at $0.18/1M = **$90**. The premium models are 30-40% more expensive than OpenAI's flagship — but they typically beat OpenAI on retrieval benchmarks (see next section).

**Cohere embed-v4.0** at $0.12/1M × 500M = **$60 one-time**. Roughly comparable to OpenAI text-embedding-3-large. The cost-per-token isn't the differentiator — the 128k context window is.

**Re-embedding cost**: if you upgrade models or re-chunk, you pay the embedding cost again. This argues for picking a model you'll stick with for 12+ months, not the cheapest model that just shipped this week. Voyage and Cohere have been more stable on their embedding model lineup than OpenAI — text-embedding-3 launched in early 2024 and has held, but earlier models (ada-002) were deprecated and forced re-embeds.

**Verdict on cost**: at typical RAG scales (single-digit-million documents), embedding cost is a one-time spend in the $10-100 range. Optimize on quality, not cost. Reserve cost optimization for the operational LLM-inference layer, which is 100-1000x more expensive over the lifetime of the system.

The MTEB (Massive Text Embedding Benchmark) leaderboard is the industry-standard public evaluation for embedding models — 56 tasks across 8 categories (retrieval, reranking, clustering, classification, semantic similarity, summarization, pair classification, bitext mining). Higher score = better, on a 0-100 scale.

**Top public scores as of June 2026** (full-MTEB average): voyage-3-large ~70.1, text-embedding-3-large ~64.6, Cohere embed-v4.0 ~67.8, voyage-3 ~66.5, text-embedding-3-small ~62.3, voyage-3-lite ~60.8. The newest open-source models (e5-mistral-7b-instruct fine-tunes, NV-Embed-v2) have pushed past 72 — but cost-per-token in production is comparable to the smaller proprietary models above.

**MTEB-Retrieval specifically** (the category that matters most for RAG): voyage-3-large leads at ~62-63, Cohere embed-v4.0 ~60-61, text-embedding-3-large ~58-59, voyage-3 ~57, text-embedding-3-small ~55. The Voyage retrieval advantage is consistent across BEIR (Benchmarking IR) sub-tasks.

**Caveat: MTEB tells you about average performance across diverse tasks; your specific corpus may behave differently.** A model that scores 70 on MTEB-average could underperform a model that scores 65 on your specific code-search or domain-specific corpus. The right benchmark is your own retrieval recall@10 on a held-out set of (query, relevant-doc) pairs.

**Domain-specific benchmarks**: voyage-code-3 beats general-purpose models by 5-10 points on CodeSearchNet and HumanEval-retrieval. voyage-finance-2 beats general-purpose models by 4-8 points on financial-document retrieval. voyage-law-2 similar advantage on legal-document retrieval. If your corpus is in one of these verticals, the domain-specific model is usually worth the price.

**Verdict on benchmarks**: voyage-3-large wins MTEB-average and MTEB-Retrieval. Cohere embed-v4.0 is competitive and adds the 128k context window. OpenAI text-embedding-3-large is the most popular default and 'good enough' for most general RAG, but no longer the highest-quality option.

The number that matters more than embedding $/1M for large-scale RAG is **vector storage cost at scale**. Embedding once is $10-100. Storing 1B vectors at high dimensions in a managed vector DB is $50k-500k/year.

**Storage math**: 4-byte float32 per dimension. text-embedding-3-large at 3072 dims = 12.3 KB per vector. 1B vectors = 12.3 TB raw. Voyage-3-large at 1024 dims = 4.1 KB per vector. 1B vectors = 4.1 TB. Cohere embed-v4.0 at 1024 dims = same 4.1 TB.

**Managed vector DB pricing** (Pinecone, Weaviate Cloud, Qdrant Cloud, Vespa Cloud) is in the $30-100/GB-month range for serverless tiers, plus query cost. 12.3 TB on Pinecone serverless at the standard tier = ~$60k/month. 4.1 TB on the same tier = ~$20k/month. **That $40k/month delta dwarfs any difference in embedding generation cost.**

**Matryoshka to the rescue**: OpenAI text-embedding-3-large supports Matryoshka-style dimension truncation — you can request 1024 dims (or 256) at query time and storage cost drops to 4.1 TB or 1.0 TB respectively. Quality degradation from 3072 → 1024 is typically 2-4% on MTEB-Retrieval. Cohere embed-v4.0 has similar flexible-dim support (256/512/1024/1536).

**Voyage's 1024-dim default** is the right answer for most production RAG systems — you get top-of-leaderboard retrieval quality without the storage tax that 3072-dim models impose. Designed-from-the-start lower-dim flagship.

**Verdict on dims**: at small scale (single-digit-million vectors), dimension count doesn't matter much. At large scale (100M+ vectors), pick a 1024-dim model or truncate a higher-dim one. The $/storage-month delta will outpace the $/quality-improvement delta for most use cases.

Most embedding models force you to chunk documents into pieces small enough to fit the input window. Chunking is where RAG systems get fragile — split at the wrong boundary and a query that should match a semantically coherent section instead matches two half-sections, neither of which retrieves well.

**OpenAI text-embedding-3 family**: 8,191 token input limit. Most documents over ~6,000 words need chunking. Standard chunking strategies (recursive character splitter, sentence-level with overlap, semantic chunking) all add complexity and edge cases.

**Voyage AI**: 32,000 token input limit on voyage-3 and voyage-3-large. 4x OpenAI's window — most documents (full-page legal contracts, technical specs, research papers under ~25k words) fit without chunking. Significantly reduces chunking complexity for mid-length documents.

**Cohere embed-v4.0**: 128,000 token input limit. The largest in the industry. Most full-length documents (book chapters, RFP responses, complete research papers, full contracts) fit without any chunking. This is the headline architectural advantage of Cohere — you can sidestep the chunking problem entirely for documents up to roughly 90,000 words.

**Trade-off**: embedding a 100k-token document produces a single 1024-dim vector that summarizes the entire document. Retrieval against that single vector is coarser than retrieval against 50 chunk-vectors. For some use cases (document-level classification, similarity search across full contracts) this is exactly what you want. For others (find the specific paragraph that answers my question) you still want chunking.

**Practical pattern**: hybrid — use Cohere embed-v4.0 to embed full documents at the document level for first-pass routing (which doc is relevant?), then use chunk-level embeddings on the routed documents for paragraph-level answers. The 128k window enables this two-tier retrieval architecture without adding a separate document-classification model.

**Cohere embed-v4.0** is the multilingual leader by design — explicit support for 100+ languages with strong cross-lingual retrieval (query in English, retrieve documents in Spanish or Japanese). The cross-lingual capability is the differentiator: most embedding models support multiple languages but underperform when queries and documents are in different languages.

**Voyage voyage-3-large** is officially multilingual (50+ languages) with strong English, Spanish, French, German, Portuguese, Mandarin, Japanese performance. Cross-lingual is weaker than Cohere but adequate for most multilingual RAG (where queries and docs tend to share a language).

**OpenAI text-embedding-3** is officially multilingual but training corpus is English-dominated. Performance on tier-1 European and Asian languages is solid; performance on Hindi, Arabic, Korean is mid-tier; performance on lower-resource languages (Bengali, Tamil, Vietnamese, most African languages) is noticeably weaker than Cohere's multilingual model.

**Specialized multilingual benchmark**: MMTEB (Massive Multilingual Text Embedding Benchmark) — Cohere embed-v4.0 leads with ~64 average across 100+ languages; Voyage at ~58; OpenAI at ~52.

**Verdict on multilingual**: Cohere wins for multilingual RAG, especially cross-lingual retrieval. Voyage is fine for selective multilingual coverage with English as the dominant query language. OpenAI is fine for English-first RAG with occasional European or East-Asian content; not the right choice for genuinely global multilingual products.

Embedding-based retrieval is good at recall (finding the candidate documents) but mediocre at precision (ranking them correctly). The two-stage RAG pattern — embed + retrieve top-N candidates, then rerank with a cross-encoder model — improves precision dramatically. Reranking is a separate API call, separate billing line, separate model.

**Cohere rerank-v3.5** at $1 per 1,000 queries (each query reranks up to 100 documents). Industry-leading reranker quality, multilingual, designed to pair with Cohere embed-v4.0. Cohere's go-to-market positions rerank as the missing piece most RAG systems need.

**Voyage rerank-2** at $0.05/1M tokens (different pricing model — by reranked tokens, not by query). Competitive quality, designed to pair with Voyage embeddings. Domain-specialized reranker variants (rerank-finance-2, rerank-law-2) exist.

**OpenAI does not offer a dedicated reranker product** as of June 2026. Most OpenAI-embedding RAG systems either skip reranking, use Cohere or Voyage as the reranker, or DIY with a fine-tuned cross-encoder on Hugging Face. The lack of a native OpenAI reranker is a meaningful gap.

**When reranking matters**: any RAG system where retrieved-doc precision affects user-facing answer quality (most production RAG). The cost is minor ($1/1k queries on Cohere = $0.001/query), the quality lift is typically 10-30% on retrieval-precision metrics.

**Verdict on rerank**: pair embeddings with a reranker for any serious production RAG. Cohere rerank-v3.5 is the best general-purpose option even if your embeddings are from a different vendor (cross-vendor pairings work fine).

Internal company-docs RAG — 1M documents averaging 800 tokens, mostly English, retrieval quality matters more than absolute cost. Vector DB on Pinecone serverless.

**OpenAI text-embedding-3-large at 1024 dims (Matryoshka-truncated)**: 800M tokens × $0.13/1M = $104 one-time embed. Storage: 1M × 4.1 KB = 4.1 GB on Pinecone serverless ≈ $200/month. MTEB-Retrieval ~58. Plus optional Cohere rerank-v3.5 at $1/1k queries = $1k/month at 1M queries.

**Voyage voyage-3-large**: 800M × $0.18/1M = $144 one-time embed. Same 1024 dims = $200/month storage. MTEB-Retrieval ~62-63 (best-in-class). Plus optional Voyage rerank-2 at $0.05/1M reranked tokens.

**Cohere embed-v4.0 at 1024 dims**: 800M × $0.12/1M = $96 one-time. Same storage. MTEB-Retrieval ~60-61. Plus Cohere rerank-v3.5 native pairing.

**Verdict**: general-purpose company-docs RAG → Voyage voyage-3-large for the retrieval quality advantage, or Cohere embed-v4.0 + rerank-v3.5 for the integrated reranker story. OpenAI text-embedding-3-large is the safe default but no longer the quality leader. Cost is rounding error across all three.

Code search across a 10M-snippet corpus — internal monorepo, GitHub issues, Stack Overflow archive. Queries in natural language, results are code snippets. Retrieval quality is everything (a wrong snippet wastes the developer's time).

**Voyage voyage-code-3** at $0.18/1M × 10M chunks × 200 avg tokens = $360 one-time embed. Specialized for code: beats voyage-3-large by 5-10 points on CodeSearchNet retrieval. 1024 dims = ~40 GB at this scale = ~$2k/month storage on Pinecone.

**OpenAI text-embedding-3-large**: 10M × 200 × $0.13/1M = $260 one-time. MTEB-Retrieval ~58 on general; under-performs voyage-code-3 by 5-10 points on code-specific benchmarks. Same 1024 dims storage.

**Cohere embed-v4.0**: $240 one-time. Multilingual strength is wasted on code (which is mostly ASCII). MTEB-Retrieval ~60 on general; not code-specialized.

**Verdict**: code search RAG → Voyage voyage-code-3, hands down. The specialized model is one of the strongest arguments for picking Voyage as a vendor. The $100 difference in embedding cost is invisible vs the daily developer time saved by better retrieval.

Legal-tech RAG: 100k full-length contracts averaging 25k tokens each. Need both document-level routing ('which contract is relevant?') and paragraph-level Q&A ('what does section 4.2 say about indemnification?'). Cross-document comparison queries are common.

**Cohere embed-v4.0 at 128k context**: embed each full contract as a single document-level vector (2.5B tokens × $0.12/1M = $300 one-time) plus chunk-level embeddings for paragraph Q&A. Two-tier architecture: document-level routing first, then chunk-level Q&A on routed documents. 128k window = no chunking required at document level.

**Voyage voyage-law-2 at 32k context**: each contract requires 1-2 chunks at document level (still much fewer than 8k chunking). Specialized for legal language — beats general-purpose embeddings by 4-8 points on legal-doc retrieval. $0.18/1M × 2.5B = $450 one-time.

**OpenAI text-embedding-3-large at 8k context**: every 25k-token contract needs 3-4 chunks at minimum. Loses the document-level summarization advantage. $0.13/1M × 2.5B = $325 one-time. No legal-domain specialization.

**Verdict**: long-document legal/contract RAG → Cohere embed-v4.0 for the 128k context + multilingual strength (international contracts) OR Voyage voyage-law-2 for the domain specialization. Hybrid pattern (Cohere for document routing, Voyage for chunk-level legal Q&A) is overengineered for most teams but optimal for the truly large-scale legal-tech use case.

**Mistake 1: defaulting to OpenAI because you already use their LLM.** OpenAI embeddings are fine — 'fine' meaning competitive but no longer the quality leader. Voyage voyage-3-large beats text-embedding-3-large on every public retrieval benchmark, often by meaningful margins. Default-vendor convenience is real but should be weighed against measurable retrieval-quality gains.

**Mistake 2: optimizing embedding cost while ignoring storage cost.** Embedding 1M docs once is $10-100. Storing the resulting 3072-dim vectors at scale is $200k-2M/year on managed vector DBs. Pick a 1024-dim model (Voyage default, Cohere flexible, OpenAI Matryoshka-truncated) to cut storage by 3x.

**Mistake 3: skipping the reranker.** Two-stage retrieval (embed + retrieve top-100, then rerank top-10) improves precision by 10-30%. Cohere rerank-v3.5 at $1/1k queries is the cheapest 30% retrieval quality lift you'll ever buy. Add it to your RAG pipeline even if your embeddings are from a different vendor.

**Mistake 4: locking into a model that gets deprecated.** OpenAI deprecated ada-002 in 2024 and forced re-embedding for users on the old model. Voyage and Cohere have been more model-stable. Factor re-embedding risk into your vendor choice — the embedding cost is one-time, but only if you don't have to do it again in 18 months.

**Mistake 5: not benchmarking on your own corpus.** MTEB is a useful prior but your specific retrieval task may behave differently. Build a held-out (query, relevant-doc) eval set of 100-500 pairs and measure recall@10 for each candidate model. The 30 minutes of eval work prevents months of suboptimal retrieval.

Pricing in this guide is sourced as follows. **OpenAI**: openai.com/api/pricing/, fetched 2026-06-20. text-embedding-3 launched at $0.02 (small) and $0.13 (large) in January 2024 and has held at those prices. The deprecated text-embedding-ada-002 was discontinued in 2024 forcing migrations. Batch API at 50% off launched in 2024.

**Voyage AI**: docs.voyageai.com/docs/pricing, fetched 2026-06-20. Voyage was acquired by MongoDB in late 2024 and integrated into the MongoDB Atlas Vector Search product line; the standalone Voyage API remains and pricing has held. voyage-3-large launched in 2025 with the 1024-dim flagship strategy. Domain-specialized variants (code-3, finance-2, law-2) priced at the same $0.18/1M.

**Cohere**: cohere.com/pricing, fetched 2026-06-20. embed-v4.0 launched in 2025 with the 128k context window expansion (up from 8k on embed-v3). Pricing at $0.12/1M has held. rerank-v3.5 launched in late 2025 with multilingual improvements over rerank-v3.

**MTEB benchmarks**: scores cited are from the public MTEB leaderboard (huggingface.co/spaces/mteb/leaderboard) as of June 2026. Note that the leaderboard updates continuously as new models are submitted — verify current standings before committing to a long-term embedding architecture.

**Live-verify before procurement**: open each vendor's pricing page and confirm per-million-token rates, max input context, and rerank pricing match this guide. Embedding pricing has been stable since 2024 but model lineups change — newer models may have replaced the ones cited above by the time you read this.