What is the best reranker model in 2026?

For most teams, Voyage rerank-3-lite at $0.02 per million tokens with 200 million free tokens per account — it matches the previous generation's quality at 40% of the price and needs one API call to add. If quality is the only constraint, Voyage rerank-3 at $0.05 per million tokens tops Voyage's own 95-dataset suite on every first-stage retriever tested. If you must self-host, BGE-reranker-v2-m3 (Apache 2.0) is the free default and Jina Reranker v3.5 is the stronger 0.6B option with a 131K context window.

How much does reranking cost per query?

Less than you expect. Reranking 100 candidate snippets of 200 tokens each is about 20,000 tokens, so roughly $0.001 on Voyage rerank-3 ($0.05/MTok) or $0.0004 on rerank-3-lite. Cohere bills per search instead: $0.0025 for Rerank 4 Pro, $0.002 for Rerank 4 Fast, $0.001 for Rerank v3.5, where one search is a query plus up to 100 documents. At a million queries a month you are looking at hundreds of dollars, not tens of thousands.

Is an open-source reranker good enough for production?

Often yes. BGE-reranker-v2-m3 is the most deployed open-weight reranker in production RAG, runs on a single consumer GPU, is multilingual, and costs nothing per query. The honest trade-off is out-of-domain quality: proprietary rerankers generalise better to corpora they were not trained on, so open weights usually want a domain fine-tune. Choose open weights for data residency, cost at very high volume, and air-gapped deployment.

Does reranking actually improve RAG accuracy?

It is typically the highest-return change available in a retrieval pipeline, because it fixes a structural weakness rather than tuning a parameter. Embedding search scores query and document independently, so it cannot model their interaction; a reranker scores them jointly on a short candidate list of about 100, which is affordable precisely because the list is short. It also requires no re-indexing, so it is reversible in an afternoon.

Quick Answer

Best Reranker Models 2026: Ranked by Quality and Price

Published: October 3, 2026

The ranking (October 2026)

This is a recommendation order, not a single score ordering — the right reranker depends mostly on billing model, deployment constraint and candidate length. Prices verified October 3, 2026.

#	Model	Price	Context	Deploy	Pick it when
1	Voyage rerank-3-lite	$0.02 / MTok	32K	API	Default. Best value, 200M free tokens
2	Voyage rerank-3	$0.05 / MTok	32K	API	Max quality, long documents, code
3	Cohere Rerank 4 Fast	$0.002 / search	32K	API	Flat cost, long candidates, 100+ languages
4	BGE-reranker-v2-m3	Free (Apache 2.0)	~8K	Self-host	Zero per-query cost, data residency
5	Jina Reranker v3.5	Token packages	131K	API + self-host	Huge candidates, air-gapped, low latency
6	Cohere Rerank 4 Pro	$0.0025 / search	32K	API	Enterprise support, max Cohere accuracy
7	Qwen3-Reranker 8B	Free (open weights)	32K	Self-host	Open weights with more headroom, multilingual
8	mxbai-rerank-large-v2	Free (Apache 2.0)	~8K	Self-host	Latency-sensitive self-hosted reranking

1. Voyage rerank-3-lite — the default

$0.02 per million tokens, 32K context, 200 million free tokens per account.

Voyage states that rerank-3-lite matches the retrieval quality of the previous-generation rerank-2.5 at 40% of the price, and beats Cohere Rerank v4.0 Pro by 2.23% NDCG@10 across Voyage’s 95-dataset suite. The 200-million-token free grant means you can validate reranking on your actual corpus for literally nothing before committing budget. Start here.

2. Voyage rerank-3 — the quality pick

$0.05 per million tokens, 32K context, instruction-following.

Released September 30, 2026. On Voyage’s suite it is the top performer on every first-stage retriever tested (BM25, OpenAI text-embedding-3-large, voyage-3-large, voyage-4-large), beating Cohere Rerank v4.0 Pro by 2.72% overall and by 13.84% on long documents (LongEmbed: NarrativeQA, SummScreenFD, QMSum). Code retrieval improved 2.17% over rerank-2.5 — Voyage explicitly retrained for coding-agent queries, which now make up a large share of its traffic.

Two practical virtues beyond score: it is a drop-in replacement for rerank-2.5 (same API, same price, same context), and relevance scores are calibrated to match rerank-2.5’s distribution, so existing score thresholds keep working. That is the difference between a one-line swap and a week of re-tuning.

Caveat: these are vendor-published numbers, with Voyage’s own embedder as the headline first stage and Qwen3-Reranker-8B’s cost assumed rather than measured. A 2.72% margin is real but modest. Test the 13.84% long-document claim yourself if long candidates are your workload — that is the one that would change an architecture.

3. Cohere Rerank 4 Fast — the flat-rate choice

$0.002 per search, 32K context, 100+ languages.

Cohere bills per search — one query plus up to 100 documents, regardless of document length. This inverts the cost maths against token-billed rivals. For 100 long documents at 8K tokens each (~800K tokens), Voyage rerank-3 costs about $0.04 while Cohere charges a flat $0.002, roughly 20× cheaper. For 100 short 200-token snippets, Voyage wins. Know your candidate length distribution before choosing a billing model — this is the single biggest cost lever in reranking and almost nobody checks it.

Rerank 4’s 32K context (quadrupled from v3.5’s 4,096 tokens) is what makes long-candidate reranking viable at all on Cohere. Rerank 4 Pro at $0.0025 buys maximum accuracy and deeper reasoning; Rerank v3.5 remains at $0.001 if 4K context is enough.

4. BGE-reranker-v2-m3 — the free default

Apache 2.0, ~568M parameters, multilingual, single consumer GPU.

Still the most widely deployed open-weight reranker in production RAG, and the correct answer when per-query fees or data residency are hard constraints. Sub-40ms p50 latency is achievable on modest hardware. The trade-off is out-of-domain generalisation: expect to fine-tune on your domain to match proprietary quality, and budget engineering time accordingly. Free per query is not free in total.

5. Jina Reranker v3.5 — the long-context and self-host option

0.6B parameters, 131K context, listwise, released July 27, 2026.

The standout spec is the 131K context window — 4× Voyage and Cohere — which matters when you rerank genuinely large documents rather than chunks. Its listwise architecture scores multiple candidates concurrently instead of pair-by-pair, which is where its latency advantage comes from (Jina reports up to 56% faster than v3 on long documents). Reported strength in legal and structured-data retrieval. Available via the Jina API, Hugging Face weights, Elastic’s inference integration, and air-gapped deployment. Check the commercial licence for production use of the weights.

6–8. The rest, briefly

Cohere Rerank 4 Pro ($0.0025/search) — the accuracy-first Cohere tier, with the strongest enterprise support story in the category. Qwen3-Reranker (0.6B / 4B / 8B, open weights) — strong multilingual accuracy across 100+ languages, 32K context on the 8B, a credible open alternative when BGE is not enough. mxbai-rerank-large-v2 (Apache 2.0) — RL-tuned, permissively licensed, a good latency-sensitive self-hosted pick.

The decision in four questions

Can you call an external API? No → BGE-reranker-v2-m3, Jina v3.5 self-hosted, or Qwen3-Reranker.
Are your candidates long (>2K tokens each)? Yes → Cohere’s per-search billing, or Jina v3.5 for very long candidates. No → Voyage’s per-token billing is cheaper.
Do you need 100+ languages with vendor support? Yes → Cohere Rerank 4.
Otherwise → Voyage rerank-3-lite, upgrade to rerank-3 if evaluation shows the quality is worth 2.5×.

What not to do

Do not use a frontier LLM as your reranker. Asking a $2/$10-per-MTok model to rank 100 candidates costs orders of magnitude more than a $0.05/MTok dedicated reranker, adds seconds of latency, and returns no calibrated score you can threshold on. Dedicated rerankers exist because this task is narrow enough to solve at 0.6B parameters. Reserve LLM ranking for when you need the ranking explained in prose.

Do not skip the baseline measurement. Reranking helps most when your first-stage retrieval has decent recall but poor precision at top-10. If recall@100 is already bad, a reranker cannot invent documents the retriever never returned — fix retrieval first.

Last verified: October 3, 2026. Prices from vendor pricing pages; comparative scores are Voyage AI’s published benchmark results.