Voyage rerank-3

Hosted API · Voyage AI · Updated 5 Oct 2026

Voyage AI’s current rerankers are rerank-3, which Voyage call their highest-accuracy model and recommend for most applications, and rerank-3-lite, tuned for latency. Both accept 32,000 tokens of query plus document, and both are billed per token with 200 million free tokens per account. The previous rerank-2.5 generation is still served at the same price — and it’s the one Voyage document for natural-language instructions and the Batch API.

Available models

Per request the query can be up to 8,000 tokens, the query plus any one document up to 32,000, with up to 1,000 documents and 600K tokens in total — counting the query once per document.

Instructions — a line such as “prefer passages that cite a statute” added to the query — reshape relevance without training data. Voyage introduced them with rerank-2.5, and their rerank-3 announcement says the Rerank 3 series keeps them: on the MAIR instruction-following benchmark, rerank-3 scores on par with rerank-2.5 in Voyage’s own test.

What changed in rerank-3

Voyage released rerank-3 and rerank-3-lite on 30 September 2026 as a drop-in upgrade to the 2.5 series: same API, context length, instruction support and price, and relevance scores calibrated to match the 2.5 models, so thresholds you tuned on 2.5 should still hold. rerank-2.5 remains available.

• Largest gains on long documents and code. On LongEmbed, rerank-3 beats rerank-2.5 by 3.35%; on Voyage’s code datasets the gain is about 2%.
• Against other rerankers. Across Voyage’s 95-dataset suite (top 100 from four first-stage retrievers, NDCG@10), Voyage report rerank-3 ahead of Cohere Rerank v4.0 Pro by 2.72% and of Qwen3-Reranker-8B by 3.02%.
• rerank-3-lite matches rerank-2.5’s quality at 40% of its price, per Voyage.

These are Voyage’s own evaluations; no independent benchmark covers rerank-3 yet. Source: Voyage’s announcement. Voyage also publish each model’s tokenizer — no weights — on Hugging Face, which you can use to count tokens locally before you’re billed for them.

Pricing

Billed tokens = query tokens × number of documents + all document tokens. At list price, 100 documents of 500 tokens each (query included) cost $0.0025 on rerank-3 — Voyage’s own example, and exactly the price of one Cohere Rerank 4 Pro search. Voyage’s pricing page still says in one sentence that rerank-2.5 includes free tokens, while its price table lists rerank-2.5 under older models with none; we follow the table. Prices from docs.voyageai.com/docs/pricing, verified 1 October 2026 and re-checked daily by this site’s source check.

Quick start

Python

pip install voyageai

import voyageai

vo = voyageai.Client(api_key="YOUR_API_KEY")

query = "How do I add reranking to my RAG pipeline?"
documents = [
    "Rerankers score each query-passage pair 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.",
]

result = vo.rerank(
    query=query,
    documents=documents,
    model="rerank-3",
    top_k=3,
)

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

REST (curl)

curl https://api.voyageai.com/v1/rerank \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "rerank-3",
    "query": "How do I add reranking to my RAG pipeline?",
    "documents": ["Rerankers score...", "BM25 is..."],
    "top_k": 3
  }'

In a RAG pipeline

import voyageai

vo = voyageai.Client(api_key="YOUR_API_KEY")

def rag_answer(query: str, vector_db, llm) -> str:
    # Stage 1: retrieve wide
    candidates = vector_db.search(query, top_k=50)
    # Stage 2: rerank tight
    result = vo.rerank(query=query, documents=candidates, model="rerank-3", top_k=5)
    top5 = [candidates[r.index] for r in result.results]
    # Stage 3: generate
    return llm.complete(f"Context:\n" + "\n\n".join(top5) + f"\n\nQ: {query}")

Pros and cons

Other models