Cohere Rerank 5

Hosted API · Cohere ·

Hosted APIMultilingualFree trial key

Cohere Rerank was one of the first commercial reranking APIs and remains among the most widely deployed. The current generation, Rerank 5, was released on 21 October 2026 in two variants: rerank-v5.0-pro (about 2B parameters) for quality and rerank-v5.0-fast (about 500M) for throughput. Both take 32,768 tokens and 100+ languages, return calibrated scores, and can point to the passages that matter inside each document. Rerank 4 remains available.

Available models

Model IDLanguagesContextNotes
rerank-v5.0-pro100+ langs32,768 tokensCurrent flagship, about 2B parameters; best for reasoning-heavy queries and long or domain-specific documents
rerank-v5.0-fast100+ langs32,768 tokensAbout 500M parameters; for high-volume search and agent loops
rerank-v4.0-pro100+ langs32,768 tokensPrevious flagship (April 2026)
rerank-v4.0-fast100+ langs32,768 tokensLower latency, higher throughput
rerank-v3.5100+ langs4096 tokens/docPrevious multilingual flagship

For new work, start with rerank-v5.0-pro, or rerank-v5.0-fast when throughput or tail latency matters more; Cohere say Rerank 5 Fast outperforms Rerank 4 Pro. Upgrading is a change to the model parameter: a reranker reads raw text at query time, so nothing needs re-indexing. Rerank 5 is billed per token, at $0.05 (Pro) or $0.02 (Fast) per million tokens, rather than per search like Rerank 4: see pricing.

What’s new in Rerank 5

"relevant_spans": {"mode": "enabled", "threshold": 0.3, "max_spans": 10}

A higher threshold keeps less text, and max_spans caps the passages per document. Cohere note that Context Focus suits targeted questions; summarisation needs the whole document, and tables, forms and code split less cleanly than prose. It is available through Cohere’s own API.

What it doesn’t do: Rerank 5 takes text only, and it doesn’t accept instructions that change how relevance is judged. For that, see instruction-following rerankers such as Voyage rerank-3 or Qwen3-Reranker.

Where it runs: the Cohere API, Cohere’s Model Vault, Microsoft Foundry and Amazon SageMaker, or a private deployment. Amazon Bedrock’s rerank API still lists only Rerank 3.5.

Pricing

TierPriceLimits
rerank-v5.0-pro$0.05 / 1M tokensBilled per token, not per search; Cohere don’t say how the query is counted
rerank-v5.0-fast$0.02 / 1M tokensSame billing unit
Trial key$01,000 API calls a month; Rerank limited to 10 requests a minute; not for production or commercial use
rerank-v4.0-pro$0.0025 / search (last listed)A search = 1 query + up to 100 documents; a document over 500 tokens (query included) counts as one per 500-token chunk
rerank-v4.0-fast$0.002 / search (last listed)Same billing unit
Model VaultFrom $3.00 per GPU-hourDedicated instances, billed by GPU type and number of replicas rather than by model; the figure is the on-demand rate for an L4
Private deploymentCustomOn-premises or isolated VPC

Prices from cohere.com/pricing and Cohere’s rate-limit docs, re-checked daily by this site’s source check; Cohere’s Rerank 5 launch post gives the same per-token prices. When Cohere priced Rerank 5 on 21 October 2026, its pricing page stopped listing Rerank 4, which is still served. The Rerank 4 prices above ($2.50 and $2 per 1,000 searches) are the last it listed, verified daily from 1 to 21 October. The 500-token rule applies to those per-search prices: see the cost calculator for what it does to a real workload.

Through a cloud instead: Amazon Bedrock sells Cohere Rerank 3.5 at $2.00 per 1,000 queries of up to 100 documents. Pinecone serves cohere-rerank-4-fast inside its vector database, and since 31 August 2026 has answered requests for cohere-rerank-3.5 with 4-fast.

Quick start

Python

pip install cohere
import cohere

co = cohere.ClientV2("YOUR_API_KEY")

query = "How do I add reranking to my RAG pipeline?"
documents = [
    "Rerankers score each query-passage pair jointly with a cross-encoder.",
    "BM25 is a classical keyword-based retrieval method.",
    "London is the capital of the United Kingdom.",
    "Two-stage retrieval: retrieve 50 candidates, rerank to top 5.",
]

results = co.rerank(
    model="rerank-v5.0-pro",
    query=query,
    documents=documents,
    top_n=3,
)

for r in results.results:
    print(f"{r.relevance_score:.4f}  {documents[r.index][:80]}")

Node.js

import Cohere from "cohere-ai";
const co = new Cohere.Client({ token: "YOUR_API_KEY" });

const result = await co.rerank({
  model: "rerank-v5.0-pro",
  query: "How do I add reranking to my RAG pipeline?",
  documents: ["...", "..."],
  topN: 3,
});

result.results.forEach((r) => console.log(r.relevanceScore, documents[r.index]));

REST (curl)

curl https://api.cohere.com/v2/rerank \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "rerank-v5.0-pro",
    "query": "How do I add reranking to my RAG pipeline?",
    "documents": ["Rerankers score query-passage pairs...", "BM25 is..."],
    "top_n": 3
  }'

Pros and cons

Pros

  • Mature, production-grade API with an SLA
  • Consistent quality across 100+ languages
  • SDK support: Python, Node, Java, Go, curl
  • Free trial key for prototyping (1,000 calls a month)
  • 32,768-token context window
  • Rerank 5: calibrated scores and Context Focus
  • Rerank 5 bills per token, so short passages cost less

Cons

  • Closed weights — you depend on Cohere
  • Rerank 4’s per-search billing counts passages over 500 tokens as several documents
  • Self-hosting means a private-deployment contract
  • No task instructions, even in Rerank 5
  • Every call is a network round trip — measure from your own region

Try reranking without an API key

Our demo runs a cross-encoder entirely in your browser — no signup, no cost.

Open the demo →

Other models

Deciding between them? Compare the rerankers on the same benchmarks →