Cohere Embed 5 is Cohere's new pair of embedding models for enterprise search, RAG, and agent retrieval. The company released Embed 5 Pro and Embed 5 Fast on September 30, 2026. Both models share one vector space, so India and USA agencies can index for quality and query for speed without rebuilding the index.

I read Cohere's announcement, then cross-checked RuntimeWire and DataNorth coverage the same day. This note is for delivery leads who already run RAG for clients and need a clear migration brief, not another vendor summary deck.

If your retainers include knowledge search, support copilots, or finance document Q&A, Cohere Embed 5 changes the cost sheet. It does not erase reindexing work, evaluation time, or the need for human QA on client corpora.

Cohere Embed 5 Pro and Fast product artwork from Cohere announcement — Cohere Embed 5 launch artwork from Cohere's September 30, 2026 blog post. Image credit: Cohere.

What Cohere Embed 5 Actually Ships

Embedding models turn text or images into vectors so search can find meaning, not just keywords. Most agencies still pick one model and use it for both indexing and live queries. That forces a single compromise between quality, latency, and spend.

According to the Cohere Embed 5 announcement, Embed 5 Pro targets maximum retrieval quality across multimodal, multilingual, financial, code, and parsed-document workloads. Embed 5 Fast targets latency and cost-sensitive query paths while keeping the same 128K-token context window and language coverage.

Both tiers accept text, images, and fused text-image inputs. Both support Matryoshka dimensions from 2048 down to 256, plus float, int8, and binary outputs. Cohere lists availability on the Cohere API, Model Vault, Microsoft Foundry, and Amazon SageMaker, with private serving through vLLM.

Pricing and access facts to put in the brief

Embed 5 Pro: $0.12 per million text tokens; $0.40 per million image tokens
Embed 5 Fast: $0.08 per million text tokens; $0.40 per million image tokens
Context: 128K tokens on both tiers
Languages: more than 100, with Hindi among the languages Cohere highlights for large gains over Embed 4
Integrations: LangChain, Haystack, Weaviate, Qdrant, Pinecone, Elasticsearch, MongoDB, Redis, Milvus, and OpenSearch

RuntimeWire's Embed 5 report stresses the operating idea: reserve Pro for indexing large collections, then run Fast on the repeated query path. That pattern only helps if your query volume dwarfs reindex volume. Measure both before you rewrite a retainer.

Why the Shared Embedding Space Matters for Agencies

The headline feature is not a single benchmark number. It is one shared embedding space for Pro and Fast. Cohere says vectors from either model can be compared directly.

In Cohere's tests across 40 development datasets, indexing with Pro and querying with Fast scored 98.4 when Pro-on-Pro was normalized to 100. Fast-on-Fast scored 96.6. Cohere also says average cross-model losses were about 1.6% and 2.7% depending on the mix, with no dataset showing a major failure in that internal set.

DataNorth's Embed 5 analysis makes the practical point well: 1.6 points on a company reranking suite is not your client's production gap. Run both setups on real queries for a week before you promise savings in a Statement of Work.

Cohere Embed 5 ViDoRe V3 retrieval quality chart comparing Pro and Fast — ViDoRe V3 retrieval quality chart from Cohere's Embed 5 announcement. Image credit: Cohere.

Recommended default pattern

Index with Pro when document quality and layout matter more than ingest speed.
Query with Fast for interactive search, support bots, and agent loops that fire many retrievals per task.
Keep output dimension and quantization identical on both sides so compatibility holds.
Use int8 at 1024 dimensions as Cohere's suggested balance for most deployments.
Reserve binary vectors for a cheap first-pass stage before a higher-precision rerank.

Benchmarks Agencies Should Quote Carefully

Cohere reports Embed 5 Pro at 85.8 on ViDoRe V3, an 8.8 point gain over Embed 4. That sits ahead of Voyage 4 Large at 83.7, Gemini Embedding 2 at 83.2, and OpenAI text-embedding-3-large at 75.5 in Cohere's comparison set.

On finance-focused tests, Cohere says Pro ranked first on FinanceBench (80.1), FinQA (90.0), and ViDoRe V3 Finance (85.0), with Fast second on each. Pro's average lead over the next non-Cohere model, Gemini Embedding 2, was 3.3 points across those finance sets. Versus OpenAI text-embedding-3-large, Cohere cites a 21.4 point lead on FinanceBench.

Treat every figure as company-reported until you reproduce it. Cohere also introduced RCP-nDCG@10, which grades candidates against query-specific relevance criteria rather than only fixed labels. RuntimeWire notes that Cohere used this method in its own evaluations, so competitor comparisons need that context in client decks.

Multimodal and parsed PDF reality

Many agency RAG stacks still parse PDFs to text, then embed the text. Tables lose structure. Charts disappear. Multi-column layouts scramble. Cohere positions Embed 5 for that mess: parsed-document suites, page-image embeddings, and fused text-image vectors.

On Cohere's parsed-document suite, Pro averaged 84.8 versus Voyage 4 Large at 83.6 and Gemini Embedding 2 at 80.8. On fused text-image corpora, Pro averaged 82.3 against Gemini Embedding 2 at 61.3. If your India D2C or USA enterprise clients ship scanned manuals and slide decks, those modes matter more than clean Wikipedia scores.

Cohere Embed 5 vector storage efficiency chart for Matryoshka and quantized embeddings — Vector storage efficiency chart from Cohere's Embed 5 announcement. Image credit: Cohere.

What Cohere Embed 5 Means for India and USA Delivery

Agencies win RAG work on grounded answers, predictable spend, and boring reliability. Cohere Embed 5 helps the first two if you design the stack on purpose. It does not remove data cleaning, access control, or evaluation.

India delivery pods

Language coverage: Cohere highlights large Embed 4 to Embed 5 Pro gains on Hindi (+12), Telugu (+12), and Farsi (+13). Test Hindi and regional language queries on client content, not only English gold sets.
Cost control: Fast's $0.08 per million text tokens helps high-volume support search for India-based BPO and SaaS clients.
Private deploy options: Model Vault and vLLM matter for banks, insurers, and healthcare clients that refuse shared API paths.
Staffing: Keep one retrieval engineer owning index config. Do not leave Pro/Fast mixing to junior prompt edits alone.

USA delivery pods

Finance and legal corpora: Use Pro for 10-Ks, earnings decks, contracts, and policy PDFs where layout carries meaning.
Agent loops: Fast fits copilots that issue many retrievals per user turn, the pattern Cohere calls out for agentic workflows.
Foundry and SageMaker paths: Useful when USA enterprise clients already standardize on Azure or AWS procurement.
Storage bills: Cohere's example cuts 100 million 2048-d float32 chunks from about 819 GB to 3.2 GB with 256-d binary vectors. Put that math in the infra estimate.

A Practical Migration Checklist

Do not rip out a working embedding model because a blog chart looks strong. Use this sequence with every India or USA client who asks about Cohere Embed 5 this week.

Inventory the current stack: model ID, dimension, distance metric, chunking rules, and monthly embedding spend.
Freeze a gold query set: at least 50 real user questions with accepted passages or answers.
Reindex a slice with Pro: one product line, one share drive, or one support collection.
A/B query paths: Pro-query versus Fast-query against the Pro index; keep the old model as control.
Score groundedness and cost: track answer quality, empty results, p95 latency, and dollars per 1,000 queries.
Decide dimension and precision: start at 1024-d int8 unless the client already standardized elsewhere.
Update the SOW: reindex hours, dual-run window, rollback owner, and acceptance thresholds.
Cut over in stages: search UI first, then agent tools, then batch analytics jobs.

Pricing and Packaging Shifts for Agencies

If your RAG packages still quote one embedding line item, split them. Clients will ask why indexing and querying use different models. Answer with numbers, not slogans.

Separate ingest and query estimates so Fast savings show up where they are real.
Sell an evaluation sprint before a full reindex. One week of dual-run beats a blind cutover.
Price storage redesign when moving from float32 to int8 or binary. Vector DB bills often dwarf token fees.
Keep a human review gate for finance, legal, and healthcare answers even when retrieval scores rise.
Document model IDs in tickets: embed-v5.0-pro and the Fast equivalent belong in change logs.

For broader stack choices around cheaper models and search products, see our notes on AI model distillation for agencies and Cloudflare AI Search for agency delivery. For custom architecture help, start with our software development consulting page or browse the Debugged Pro blog.

Risks and Client Talking Points

Strong vendor charts can tempt overconfident promises. Keep these limits in every India USA kickoff deck.

Benchmarks are Cohere-run; your parsing pipeline can erase the claimed gain.
RCP-nDCG@10 is a Cohere methodology focused on reranking a candidate set, not a full-corpus bake-off on your data.
DataNorth notes Cohere published throughput but not a clear millisecond latency table for Pro versus Fast.
Mixing Pro and Fast still requires matching dimensions and compatible quantization.
Reindexing large corpora has calendar cost even when per-token prices look small.
Compass Cloud, Cohere's managed search platform, is listed as private beta. Do not sell it as GA.

Plain lines that have worked on client calls:

"Cohere Embed 5 lets us index with Pro and query with Fast in one vector space."
"We will dual-run against your gold queries before we touch production search."
"Finance and scanned PDF gains look strong in Cohere tests; we still prove them on your files."
"Token price is only part of the bill. Vector storage and reindex hours matter too."
"Hindi and other language gains are interesting for India traffic, but we will score them on your content."

Sample Two-Week Pilot Plan

Use this when a USA product lead or India delivery manager asks what changes next Monday. It assumes an existing RAG app with a vector database you already operate.

Day 1-2: Access keys, current index stats, gold query export, success metrics written down.
Day 3-5: Pro reindex of a non-critical collection; wire Fast and Pro query toggles behind a flag.
Day 6-8: Side-by-side scoring on groundedness, citation hit rate, and cost per 1,000 queries.
Day 9-10: Dimension and int8 experiments; storage estimate for full corpus.
Day 11: Client review with charts from their data, not only Cohere slides.
Day 12-14: Go or no-go. If go, schedule staged cutover and name the rollback owner.

Key Takeaways

Cohere Embed 5 shipped September 30, 2026 as Pro and Fast tiers that share one embedding space.
Pro is priced at $0.12 and Fast at $0.08 per million text tokens; image tokens are $0.40 on both.
Cohere recommends indexing with Pro and querying with Fast for many production setups.
Reported gains are strongest on visually rich docs, parsed PDFs, and finance benchmarks in Cohere's tests.
India and USA agencies should dual-run on client corpora before rewriting RAG retainers.
Storage compression and reindex effort often matter as much as the per-token price cut.

FAQ

What is Cohere Embed 5?

Cohere Embed 5 is a family of embedding models released on September 30, 2026. It includes Embed 5 Pro for maximum retrieval quality and Embed 5 Fast for lower-latency, high-volume queries. Both support text, images, fused text-image inputs, more than 100 languages, and a 128K-token context window.

How much does Cohere Embed 5 cost?

Cohere prices Embed 5 Pro at $0.12 per million text tokens and Embed 5 Fast at $0.08 per million text tokens. Image inputs cost $0.40 per million tokens for either tier. The models are available on the Cohere API, Model Vault, Microsoft Foundry, and Amazon SageMaker, with private serving options through vLLM.

Can agencies index with Pro and query with Fast?

Yes. Cohere says Pro and Fast share one embedding space, so vectors from either model can be compared directly. Cohere recommends indexing with Pro and querying with Fast. In its tests across 40 datasets, that mix scored 98.4 versus 100 for Pro on both sides, when normalized.

How does Cohere Embed 5 compare on enterprise documents?

On ViDoRe V3, Cohere reports Embed 5 Pro at 85.8, ahead of Voyage 4 Large at 83.7, Gemini Embedding 2 at 83.2, and OpenAI text-embedding-3-large at 75.5. Pro also ranked first on FinanceBench, FinQA, and ViDoRe V3 Finance in Cohere's comparisons. Treat these as company-run evaluations until you retest on your corpus.

What should India and USA agency teams do this week?

Pilot Embed 5 on a slice of real client data, keep the current embedding model as a control, measure answer quality and cost separately for Pro-only versus Pro-index Fast-query, and put reindex effort plus vector dimension choices into the SOW before promising a full migration.

About the author: Nitin Kumar is Lead Project Coordinator at Debugged Pro. He works with software and IT agencies in India and the USA on delivery planning, RAG systems, and practical AI adoption inside client projects.