Cohere Rerank iconCohere Rerank

Cohere's reranking API for search and RAG: Rerank 5 Pro and Fast with 32K context, 100+ languages, Context Focus spans and calibrated scores

32K Context (Rerank 4 and 5)
100+ Languages
$0.02 Per 1M tokens (Rerank 5 Fast)

Overview

Cohere Rerank is a cross-attention reranking model: you send a query and a list of retrieved documents, and it returns them re-ordered with a relevance score, so a RAG pipeline or agent passes only the best results to the LLM. The current generation, Rerank 5 (released October 21, 2026), comes as rerank-v5.0-pro (about 2B parameters, highest quality) and rerank-v5.0-fast (about 500M, low latency). Both read 32K tokens, cover 100+ languages, and return calibrated 0 to 1 scores so one threshold works across queries. An optional Context Focus mode also returns the most relevant passages inside each document as scored character spans. Rerank reads raw text at query time, so upgrading models needs no re-indexing; structured records go in as YAML strings. Rerank 4 (Pro and Fast, 32K) and the 4K-context v3 models are still listed. It runs on the Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker, OpenRouter, North, and in private VPC or on-premises installs. Also from Cohere: Cohere Embed (embedding models) and Cohere (Command models and platform).

The Verdict

Who Should Use Cohere Rerank?

Best For

  • Adding a quality boost on top of any existing vector or keyword search
  • RAG and agent pipelines that want to cut the context sent to the LLM
  • Multilingual search across 100+ languages
  • Enterprises that need private VPC or on-premises deployment

Not Ideal For

  • Teams that want open weights to self-host for free
  • Testing at volume on a free key: trial keys allow 1,000 calls a month
  • Image-only search: Rerank 5 accepts text only

What's Great

  • Drop-in: upgrading is a change of the model parameter, no re-embedding
  • Context Focus returns the relevant passages, not just a document score
  • Calibrated scores make a fixed relevance cutoff practical
  • Cohere reports Rerank 5 Pro at 82.2 on BRIGHT, 5.8 points ahead of Voyage Rerank 3 (vendor figure)
  • Fast tier gives about 2.9x Pro's throughput on one A100, per Cohere

Watch Out For

  • Closed model: no weights to download
  • Benchmark numbers come from Cohere itself, not independent tests
  • Trial keys are rate limited (10 per minute) and not allowed for production
  • Pricing page still describes per-search units (one query, up to 100 documents) while Rerank 5 is priced per token, so check which applies to your model
  • Text only: tables and JSON must be converted to YAML strings

Pricing

View all features & details

Models

  • rerank-v5.0-pro (about 2B) and rerank-v5.0-fast (about 500M), 32K
  • rerank-v4.0-pro and rerank-v4.0-fast, 32K, semi-structured JSON
  • rerank-v3.5, rerank-english-v3.0, rerank-multilingual-v3.0, 4K

Rerank 5 Features

  • Context Focus (off by default): scored passage spans per document
  • Calibrated relevance scores from 0 to 1
  • Long documents are split and scored in several passes
  • Works on emails, tables, JSON (as YAML) and code

Where It Runs

  • Cohere API and Model Vault
  • Microsoft Foundry and Amazon SageMaker (Pro and Fast)
  • OpenRouter and Vercel AI Gateway
  • Amazon Bedrock lists Rerank 3.5 at $2 per 1,000 queries
  • Private VPC and on-premises

Real-World Usage

Customers

  • Logos on the Rerank page include Oracle, Dell Technologies, RBC, Fujitsu, SAP, Salesforce, TD Bank and McKinsey
  • Nebius is quoted praising the models for quality and latency

Context Focus Results (vendor tests)

  • Across six deep-search benchmarks, tokens fell 41% while accuracy rose from 39% to 52%
  • Dropping half the retrieved context kept about 96% of answer quality

How It Compares

Feature Cohere Rerank 5 Voyage rerank-3 jina-reranker-v3 bge-reranker-v2-m3
Context 32K tokens 32K tokens See model card See model card
Price per 1M tokens $0.02 (Fast) to $0.05 (Pro) $0.02 (lite) to $0.05 Jina token top-ups Free to self-host
Open weights No No Yes, CC BY-NC 4.0 Yes, Apache 2.0
Max documents per call Not stated (long docs split) 1,000 64 Self-managed
Owner Cohere MongoDB Elastic BAAI

Sources: Cohere Rerank 5 · Voyage reranker docs · Voyage pricing · jina-reranker-v3 card · bge-reranker-v2-m3 card

Is this your tool? Add a badge