Overview
Cohere Rerank 5 is an enterprise reranking model released in October 2026 in Pro and Fast tiers, joining Embed 5 and Parse 5 in the same generation of search models. It sits after an existing keyword, vector or hybrid search stage, needs no re-indexing, and reorders candidates before they reach the language model.
Both new capabilities in this generation target context cost. Context Focus returns the relevant passages inside each document along with per-span scores, so an app can send only those few sentences onward; Cohere measured across 30 question-answering datasets that removing half the retrieved text cost 2 points of answer quality and removing three quarters cost 7. Calibrated scores make relevance_score readable as a probability, so 0.7 means a 70% chance the document is useful and one threshold works across domains instead of one per model.
Pro costs $0.05 per million tokens and Fast $0.02. Beyond the Cohere API and Model Vault, both tiers are available on Microsoft Foundry and Amazon SageMaker, with private deployment in a customer VPC or on-premises.
Key Features
- Context Focus passage returns: A single call returns document-level relevance plus scored passages inside each document, so only the useful sentences reach the model
- Calibrated relevance scores: relevance_score reads as the probability a document is useful, where 0.7 means 70%, letting one threshold serve every query and domain
- Pro and Fast tiers: Pro targets quality-first retrieval and Fast targets latency and cost sensitive workloads, behind the same interface
- Multilingual retrieval: Cohere evaluated 20 languages across 74 datasets, with gains in Spanish, French, German, Arabic and Korean
- No re-indexing: Runs after existing BM25, vector or hybrid search and works with LangChain, Haystack, Elasticsearch and Weaviate
- Multiple deployment paths: Available through the Cohere API, Model Vault, Microsoft Foundry and Amazon SageMaker, or in a customer VPC and on-premises
Use Cases
- Enterprise knowledge bases and document Q&A
- Retrieval augmentation over contracts, filings and research
- Locating answers inside support knowledge bases
- Relevance ranking for multilingual site search
- The retrieval toolchain behind an agent
- Self-hosted retrieval systems
Pros
- Passage-level returns cut a large share of context tokens, which shows up most on long documents
- Scores work as probabilities, so thresholds are tuned once
- Two tiers cover quality-first and cost-first workloads by swapping a model name
- No re-indexing is needed, since it plugs into the retrieval stack already running
- Private VPC and on-premises deployment are supported
- Evaluation spans 20 languages, so one model serves cross-border content
Pricing
Billing is per token: $0.05 per million tokens for Rerank 5 Pro and $0.02 per million for Rerank 5 Fast. Both tiers run on the Cohere API, Model Vault, Microsoft Foundry and Amazon SageMaker, with private deployment quoted by contract and current rates listed on the official site.
Summary
Rerank 5 turns reranking into a line item you can shrink: Context Focus sends only the relevant passages onward, and calibrated scores let a single threshold travel across domains. Teams already running RAG can swap the model name without re-indexing, and on long documents the passage-level returns save more context than moving to a smaller generation model would.
Version History
- Rerank 5 ships with passage-level Context Focus and calibrated scores (2026-10-21): Cohere released Rerank 5 in October 2026 in rerank-v5.0-pro and rerank-v5.0-fast tiers, alongside Embed 5 and Parse 5. Context Focus returns document-level scores plus scored spans inside each document in one call, and relevance_score is now a calibrated probability where 0.7 means a 70% chance of being useful. Across 30 question-answering datasets Cohere measured a 2 point quality loss when half the retrieved text was removed. Pricing is $0.05 per million tokens for Pro and $0.02 for Fast.