Cohere Embed 5 Pro/Fast: shared-space GA embeddings for RAG

Cohere released Embed 5 on 2026-09-30 as a generally available Pro/Fast embedding family for enterprise search, RAG, and agentic retrieval—embed-v5.0-pro and embed-v5.0-fast—with a shared embedding space so teams can index with Pro and query with either model without re-indexing (Embed 5 blog).

This is a Desk Bot dev/rag briefing locked to that primary. Every bench and competitor score below is Cohere-reported, not desk-verified. Compass Cloud (private beta) is out of scope here.

What shipped

	Embed 5 Pro	Embed 5 Fast
Model ID	embed-v5.0-pro	embed-v5.0-fast
Context	128K tokens	128K tokens
Inputs	Text, images, fused text+image	Same
Languages	100+	100+
Output dims (Matryoshka)	2048 / 1536 / 1024 / 768 / 512 / 256	Same
Formats	float / int8 / binary	Same
Text price (blog)	$0.12 / 1M tokens	$0.08 / 1M tokens
Image price (blog)	$0.40 / 1M tokens	$0.40 / 1M tokens

Hosts Cohere names: Cohere API / Model Vault, Microsoft Foundry, Amazon SageMaker; self-host via vLLM also stated. Integrations listed: LangChain, Haystack, Weaviate, Qdrant, Pinecone, Elasticsearch, MongoDB, Redis, Milvus, OpenSearch (Embed 5 blog).

Pricing above is blog-only—do not invent other tiers. Both sides of a Pro/Fast mix must use the same output_dimension; Cohere says Matryoshka truncation and int8 stay compatible (Embed 5 blog).

Shared space (practitioner hook)

Pro and Fast share one embedding space. Cohere’s recommended pattern: index Pro → query Fast for much of all-Pro quality with Fast latency/cost on the request path. Cross-model pairs, per Cohere’s 40-dataset mean (normalized Pro+Pro=100): Fast corpus / Fast query 96.6; Pro corpus / Fast query 98.4; Fast corpus / Pro query 97.3—framed as ~1.6% / 2.7% average losses for Fast vs Pro queries respectively, with no major failure called out (Embed 5 blog).

API snippets use distinct input_type search_document vs search_query (Embed 5 blog).

Compression

Matryoshka + int8/binary cut vector storage. Cohere’s illustration: 2048-d float32 ≈ 8 KB → 1024-d int8 1 KB → 256-d binary 32 bytes; across 100M chunks, raw storage ~819 GB → 3.2 GB. Vendor default efficiency point: 1024-d int8. Fast document throughput is self-reported ~2.4× Pro across ~200-token and ~1K-token contexts (Embed 5 blog).

Benches (soft — Cohere)

Embed 5 is evaluated with RCP-nDCG@10, Cohere’s methodology: a two-stage / reorder-fixed-candidate setup whose scores reflect reranking-style quality, not a drop-in first-stage Recall substitute (Embed 5 blog footnote).

Cohere-reported highlights (attribute all):

• ViDoRe V3 (RCP-nDCG@10, parsed text): Pro 85.8 (+8.8 vs Embed 4); Fast 84.5—ahead of Voyage 4 Large 83.7, Gemini Embedding 2 83.2, OpenAI text-embedding-3-large 75.5 in Cohere’s table
• Finance: FinanceBench 80.1 Pro / 80.0 Fast; FinQA 90.0 / 88.8; ViDoRe V3 Finance 85.0 / 83.9
• Parsed-PDF suite avg: Pro 84.8 vs Voyage 4 Large 83.6, Fast 83.4, Gemini Embedding 2 80.8
• Fused text-image avg (5 sets): Pro 82.3, Fast 81.2, Gemini Embedding 2 61.3

Treat competitor rows as Cohere’s comparisons only (Embed 5 blog).

Who should care

RAG and enterprise search teams that want GA multimodal embeddings with a Pro-index / Fast-query path should start at the Embed 5 blog—lock dollars to that post, keep RCP-nDCG@10 as Cohere’s metric, and leave Compass Cloud for a separate beat.