What it is
rerank-3 and rerank-3-lite are rerankers from Voyage AI (part of MongoDB), released on 30 September 2026 to succeed the rerank-2.5 series, which stays available. A reranker takes the shortlist a search system has already retrieved and rescores each result by reading the query and the document together. That is slower than comparing vectors but more precise, so it runs on a shortlist only; Voyage's own tests rerank the top 100 results.
Both accept a query of up to 8,000 tokens, up to 32,000 tokens for the query plus any one document, and up to 1,000 documents or 600,000 total tokens per request. Voyage says they follow natural-language instructions about relevance, and their scores are calibrated to rerank-2.5, so existing cut-off thresholds keep working without code changes.
rerank-3 costs $0.05 and rerank-3-lite $0.02 per million tokens, the same as their 2.5 predecessors, with the first 200 million tokens free for new users. They run on the Voyage API and in MongoDB Atlas, including native reranking in Atlas Vector Search.
Our review
rerank-3 has the best published numbers of any hosted reranker, with the caveat that Voyage published them. Across 95 retrieval datasets, averaged over four first-stage retrievers, it reports rerank-3 ahead of Cohere Rerank v4.0 Pro by 2.72% and Qwen3-Reranker-8B by 3.02% on NDCG@10, a standard ranking-quality score. On long documents the margin over Cohere's Pro model grows to 13.84%. Cohere has published no counter-figures.
The same tables show when reranking is worth an extra network call. It lifts keyword search (BM25) by 20.33% and OpenAI's text-embedding-3-large by 16.67%, but voyage-4-large by only 6.38%. The weaker or cheaper the first stage, the more a reranker repays its latency. Voyage publishes no latency figures, so that side of the trade has to be measured in your own stack.
Per-token billing sounds cheap until you read the formula. Each request counts the query once per document plus every document's tokens, so a 50-token query over 100 passages of 500 tokens comes to 55,000 tokens. That makes rate limits the first wall rather than price: rerank-3's Tier 1 limit of 2 million tokens a minute covers about 36 such searches a minute, and Tier 3 reaches only about 109.
The data default deserves attention. Voyage stores API data and may use it for training unless an organisation admin with a payment method on file opts out, which switches to zero-day retention and cannot be undone from the dashboard. The weights are closed, so teams that must run on their own hardware are left with open models such as Qwen3-Reranker-8B, which Voyage's tests put about 3% behind.
Where it earns its place
A search engineer running hybrid BM25 and vector retrieval over 2M support articles
Raise answer quality without re-indexing
Send the top 100 candidates to rerank-3-lite; Voyage reports a 20.33% NDCG@10 gain over BM25 alone for rerank-3. At 55,000 tokens a search, 1M searches a month is about $1,100 on lite or $2,750 on rerank-3.
A legal-tech team doing RAG over long contracts
Rank whole clauses and filings, not snippets
The 32,000-token window per query and document lets long passages go in unchunked, and Voyage's long-document test shows a 13.84% lead over Cohere Rerank v4.0 Pro. Long passages multiply token counts, so cost tracks passage length rather than search count.
A MongoDB Atlas team with an existing vector search index
Add reranking without another vendor contract
Use native reranking in Atlas Vector Search; scores keep rerank-2.5's calibration, so relevance cut-offs tuned earlier still apply. Confirm the training opt-out covers your route first, since the Voyage FAQ describes it only for Voyage-hosted endpoints.
Before you commit
- Setup
- Create a Voyage API key and make one rerank call with query, documents, model and optional top_k; moving from rerank-2.5 needs only the model name changed.
- Integrations
- Voyage REST API and Python client, the MongoDB Atlas Embedding and Reranking API, and native reranking in Atlas Vector Search. AWS or Azure marketplace listings are not mentioned in the launch post.
- Data and privacy
- Voyage-hosted endpoints store data and may use it for training by default; an admin with a payment method can opt out for zero-day retention, and reversing that needs an email to Voyage's legal team. SOC 2 or HIPAA status is not published in the docs we read.
- Cost at scale
- A search with a 50-token query and 100 documents of 500 tokens counts as 55,000 tokens, so 1M searches a month is about $2,750 on rerank-3 or $1,100 on rerank-3-lite. The 200M free tokens cover only about 3,600 such searches.
- Limits
- 8,000-token query, 32,000 tokens for query plus one document, 1,000 documents and 600,000 tokens per request. Tier 1 allows 2M tokens a minute for rerank-3 and 4M for lite, tripling at Tier 3 ($1,000 paid).
What's good
- Ahead of Cohere Rerank v4.0 Pro by 2.72% across 95 datasets (Voyage's tests)
- 13.84% ahead of Cohere Pro on long documents
- rerank-3-lite at $0.02 per million tokens matches rerank-2.5
- Scores calibrated to 2.5, so thresholds carry over
What's not
- Benchmarks are Voyage's own
- No published latency figures
- Stores and may train on API data unless an admin opts out
- Tier 1 limit of 2M tokens a minute caps throughput
- Closed weights, no self-hosting
Best for
Teams running RAG or search in production who want the best hosted reranking quality for the money.
Alternatives
Voyage rerank-3: common questions
- Is Voyage rerank-3 worth it?
- On Voyage's own numbers rerank-3 leads every hosted rival, and the lite model costs so little that the rate limit, not the bill, decides how much reranking you do. We rate it 4.6 out of 5.
- How much does Voyage rerank-3 cost?
- rerank-3 $0.05 per 1M tokens; rerank-3-lite $0.02 per 1M tokens; first 200M tokens free.
- What are the downsides of Voyage rerank-3?
- Benchmarks are Voyage's own; no published latency figures; stores and may train on API data unless an admin opts out; tier 1 limit of 2M tokens a minute caps throughput; closed weights, no self-hosting.
- Who is Voyage rerank-3 best for?
- Teams running RAG or search in production who want the best hosted reranking quality for the money.
- How much does Voyage rerank-3 cost for a team or in production?
- A search with a 50-token query and 100 documents of 500 tokens counts as 55,000 tokens, so 1M searches a month is about $2,750 on rerank-3 or $1,100 on rerank-3-lite. The 200M free tokens cover only about 3,600 such searches.
- Is Voyage rerank-3 safe for company data?
- Voyage-hosted endpoints store data and may use it for training by default; an admin with a payment method can opt out for zero-day retention, and reversing that needs an email to Voyage's legal team. SOC 2 or HIPAA status is not published in the docs we read.
- What are the best Voyage rerank-3 alternatives?
- The closest alternatives we have reviewed are Cohere Rerank 4, Jina Reranker v3.5 and Qwen3-Embedding and Qwen3-Reranker.
Answers checked Oct 22, 2026.
Make Voyage rerank-3? Show this rating on your site
