Reranker models - verified October 18, 2026

Reranker models compared

We read the rerank docs, pricing pages and Hugging Face cards of 10 providers on October 18, 2026 and recorded 42 reranker models: 35 current and 7 older ones still served, covering 13 distinct open-weight checkpoints. Prices stay in each provider's own billing unit, so they are grouped by unit and ranked only inside a group.

Priced per 1M tokens, the lowest current rate is $0.02 per 1M tokens for rerank-3-lite (Voyage AI).
Priced per 1,000 queries, the lowest current rate is $1.00 per 1,000 queries, for Amazon Rerank 1.0 (Amazon Bedrock), semantic-ranker-default-002 (Google Cloud), semantic-ranker-default-003 (Google Cloud) and 4 more.
The largest stated context figure is 32,000 tokens (Rerank 4 Pro (Cohere), Rerank 4 Fast (Cohere), rerank-3 (Voyage AI), rerank-3-lite (Voyage AI) and 3 more), though providers scope it differently: per document, per query plus document or per record.

A reranker re-scores the candidates a first-stage search returns. Picking that first stage? Compare embedding models or read which embedding model to use.

Models: 42
Providers: 10
Open-weight checkpoints: 13
Older models: 7

Pricing by billing unit

Each table holds one billing unit, as the provider states it. Rows are sorted by price inside a table only; a price in one table is never compared with a price in another, and nothing is converted. Older models are listed separately below.

Per 1M tokens

5 models

Billed on tokens processed. Voyage counts the query once per document plus every document's tokens; Alibaba bills input tokens only.

Per 1M tokens: 5 reranker models
Model	Price (provider's unit)	Billing note	Context (tokens)	Docs per request
rerank-3-liteVoyage AI - `rerank-3-lite`	$0.02 per 1M tokensdocs.voyageai.com	Billed on total processed tokens: (query tokens x number of documents) + tokens in all documents. Table lists 200 million free tokens.	32,000query plus any single document	1,000hard limit
rerank-3Voyage AI - `rerank-3`	$0.05 per 1M tokensdocs.voyageai.com	Billed on total processed tokens: (query tokens x number of documents) + tokens in all documents. Table lists 200 million free tokens.	32,000query plus any single document	1,000hard limit
qwen3-rerankAlibaba Cloud Model Studio - `qwen3-rerank`	$0.1 per 1M input tokensalibabacloud.com	Billed by input tokens, output free; Singapore region, International deployment scope; 1 million free tokens valid 90 days.	4,000per query or document; 120,000 per request	500hard limit
qwen3-vl-rerankAlibaba Cloud Model Studio - `qwen3-vl-rerank`	$0.1 per 1M text input tokens; $0.258 per 1M image input tokensalibabacloud.com	Listed under the China (Beijing) pricing tab.	8,000per item; 120,000 per request	100hard limit for text documents (40 images, 4 videos)
gte-rerank-v2Alibaba Cloud Model Studio - `gte-rerank-v2`	$0.115 per 1M text input tokensalibabacloud.com	Listed under the China (Beijing) pricing tab.	4,000per item; 30,000 per request	500hard limit

Per 1,000 queries

8 models

Billed per query, where one query covers up to 100 documents (Google) or 100 document chunks of up to 512 tokens (Bedrock). Larger requests count as several queries.

“A query is a single call to the reranker model that can contain up to 100 document chunks.”

Amazon Bedrock billing unit - docs.aws.amazon.com

Per 1,000 queries: 8 reranker models
Model	Price (provider's unit)	Billing note	Context (tokens)	Docs per request
Amazon Rerank 1.0Amazon Bedrock - `amazon.rerank-v1:0`	$1.00 per 1000 queriesaws.amazon.com	Bedrock defines a query as a call with up to 100 document chunks; larger requests count as multiple queries.	Not stated	1,000hard limit (sources array of the Bedrock Rerank API)
semantic-ranker-default-002Google Cloud - `semantic-ranker-default-002`	$1.00 per 1,000 queriescloud.google.com	Pay-as-you-go column; the next two columns ($0.90 and $0.80) are 1-year and 3-year savings-plan rates. A query is up to 100 documents; every further 100 documents counts as another query.	512per record (title plus content)	1,000hard limit
semantic-ranker-default-003Google Cloud - `semantic-ranker-default-003`	$1.00 per 1,000 queriescloud.google.com	Pay-as-you-go column; the next two columns ($0.90 and $0.80) are 1-year and 3-year savings-plan rates. A query is up to 100 documents; every further 100 documents counts as another query.	512per record (title plus content)	1,000hard limit
semantic-ranker-default-004Google Cloud - `semantic-ranker-default-004`	$1.00 per 1,000 queriescloud.google.com	Pay-as-you-go column; the next two columns ($0.90 and $0.80) are 1-year and 3-year savings-plan rates. A query is up to 100 documents; every further 100 documents counts as another query.	1,024per record (title plus content)	1,000hard limit
semantic-ranker-default-005Google Cloud - `semantic-ranker-default-005`	$1.00 per 1,000 queriescloud.google.com	Pay-as-you-go column; the next two columns ($0.90 and $0.80) are 1-year and 3-year savings-plan rates. A query is up to 100 documents; every further 100 documents counts as another query.	1,024per record (title plus content)	1,000hard limit
semantic-ranker-fast-004Google Cloud - `semantic-ranker-fast-004`	$1.00 per 1,000 queriescloud.google.com	Pay-as-you-go column; the next two columns ($0.90 and $0.80) are 1-year and 3-year savings-plan rates. A query is up to 100 documents; every further 100 documents counts as another query.	1,024per record (title plus content)	1,000hard limit
semantic-ranker-fast-005Google Cloud - `semantic-ranker-fast-005`	$1.00 per 1,000 queriescloud.google.com	Pay-as-you-go column; the next two columns ($0.90 and $0.80) are 1-year and 3-year savings-plan rates. A query is up to 100 documents; every further 100 documents counts as another query.	1,024per record (title plus content)	1,000hard limit
Cohere Rerank 3.5 (Bedrock)Amazon Bedrock - `cohere.rerank-v3-5:0`	$2.00 per 1,000 queriesaws.amazon.com	Billed through AWS Marketplace under the model provider; a query can contain up to 100 document chunks.	4,000context window (model card says 4K)	1,000hard limit (sources array of the Bedrock Rerank API)

Add-on per 1K Store queries

2 models

Mixedbread sells reranking only inside its Stores API: a per-query add-on on top of the semantic Store query price, not a standalone rerank endpoint.

Add-on per 1K Store queries: 2 reranker models
Model	Price (provider's unit)	Billing note	Context (tokens)	Docs per request
mxbai-rerank-large-v2Mixedbread - `mixedbread-ai/mxbai-rerank-large-v2`	+$3.50 per 1K queries (rerank add-on to $4 per 1K semantic Store queries)mixedbread.com	Reranking is sold only inside Mixedbread Store queries, not as a standalone rerank endpoint; the FAQ gives $7.50 per thousand for semantic queries with reranking.	Not stated	Not stated
mxbai-rerank-v3.1-listwiseMixedbread - `mixedbread-ai/mxbai-rerank-v3.1-listwise`	+$3.50 per 1K queries (rerank add-on to $4 per 1K semantic Store queries)mixedbread.com	Reranking is sold only inside Mixedbread Store queries, not as a standalone rerank endpoint; the FAQ gives $7.50 per thousand for semantic queries with reranking.	Not stated	Not stated

Dedicated hourly only (Cohere Model Vault)

5 models

Cohere bills Rerank in search units. The pay-as-you-go price per search was not stated in the page text we read, so only the Model Vault dedicated-instance rates are shown.

“A single search unit is defined as one query with up to 100 documents to be ranked.”

Cohere billing unit - cohere.com

Dedicated hourly only (Cohere Model Vault): 5 reranker models
Model	Pay-as-you-go per search unit	Model Vault (per instance per hour)	Context (tokens)	Docs per request
Rerank 3.5Cohere - `rerank-v3.5`	Not statedbilled per search unit; price not in the page text read	$5.00 Mediumper instance per hour (Model Vault) - cohere.com	4,096per query plus document	1,000recommended maximum, not a hard limit
Rerank 4 FastCohere - `rerank-v4.0-fast`	Not statedbilled per search unit; price not in the page text read	$5.00 Mediumper instance per hour (Model Vault) - cohere.com	32,000context length (models table lists 32k)	1,000recommended maximum, not a hard limit
Rerank 4 ProCohere - `rerank-v4.0-pro`	Not statedbilled per search unit; price not in the page text read	$5.00 Mediumper instance per hour (Model Vault) - cohere.com $10.00 Largeper instance per hour (Model Vault) - cohere.com	32,000context length (models table lists 32k)	1,000recommended maximum, not a hard limit
Rerank English 3.0Cohere - `rerank-english-v3.0`	Not statedbilled per search unit; price not in the page text read	No rate listed	4,096per query plus document	1,000recommended maximum, not a hard limit
Rerank Multilingual 3.0Cohere - `rerank-multilingual-v3.0`	Not statedbilled per search unit; price not in the page text read	No rate listed	4,096per query plus document	1,000recommended maximum, not a hard limit

Dedicated endpoint only (Together AI)

2 models

Together runs rerank models only on dedicated endpoints billed per GPU hour. There is no per-token or per-query rerank price.

Dedicated endpoint only (Together AI): 2 reranker models
Model	Per-call price	Dedicated endpoint rate	Context (tokens)	Docs per request
Llama-Rank-V1 (Together dedicated)Together AI - `Salesforce/Llama-Rank-V1`	None (dedicated only)	$5.49 NVIDIA HGX H100 on-demandper GPU per hour (dedicated endpoint) - together.ai	8,000per document (API-level statement)	Not stated
mxbai-rerank-large-v2 (Together dedicated)Together AI - `mixedbread-ai/mxbai-rerank-large-v2`	None (dedicated only)	$5.49 NVIDIA HGX H100 on-demandper GPU per hour (dedicated endpoint) - together.ai	8,000per document (API-level statement)	Not stated

Self-host (open weights or NIM containers)

13 models

No per-call price: you download the weights or run the container and pay for the hardware.

Self-host (open weights or NIM containers): 13 reranker models
Model	Weights and license	Access	Context (tokens)	Docs per request
Qwen3-Reranker-8BQwen - `Qwen/Qwen3-Reranker-8B`	Open weightsApache 2.0 - 8.19B params	Open weights	32,000context length (card says 32k)	Not stated
Qwen3-Reranker-4BQwen - `Qwen/Qwen3-Reranker-4B`	Open weightsApache 2.0 - 4.02B params	Open weights	32,000context length (card says 32k)	Not stated
bge-reranker-v2-minicpm-layerwiseBAAI - `BAAI/bge-reranker-v2-minicpm-layerwise`	Open weightsApache 2.0 - 2.72B params	Open weights	Not stated	Not stated
bge-reranker-v2-gemmaBAAI - `BAAI/bge-reranker-v2-gemma`	Open weightsApache 2.0 - 2.51B params	Open weights	Not stated	Not stated
Llama Nemotron Rerank VL 1B v2NVIDIA NIM - `nvidia/llama-nemotron-rerank-vl-1b-v2`	Open weightsNVIDIA Open Model License - 1.68B params	Self-hosted NIM and open weights	8,192optimized NIM profiles	Not stated
Llama Nemotron Rerank 1B v2NVIDIA NIM - `nvidia/llama-nemotron-rerank-1b-v2`	Open weightsOpenMDW-1.1 - 1.24B params	Self-hosted NIM and open weights	8,192query plus passage, optimized NIM profiles	Not stated
Qwen3-Reranker-0.6BQwen - `Qwen/Qwen3-Reranker-0.6B`	Open weightsApache 2.0 - 596M params	Open weights	32,000context length (card says 32k)	Not stated
bge-reranker-v2-m3BAAI - `BAAI/bge-reranker-v2-m3`	Open weightsApache 2.0 - 568M params	Open weights	Not stated	Not stated
bge-reranker-largeBAAI - `BAAI/bge-reranker-large`	Open weightsMIT - 560M params	Open weights	Not stated	Not stated
mxbai-rerank-base-v2Mixedbread - `mixedbread-ai/mxbai-rerank-base-v2`	Open weightsApache 2.0 - 494M params	Open weights	Not stated	Not stated
mxbai-rerank-large-v1Mixedbread - `mixedbread-ai/mxbai-rerank-large-v1`	Open weightsApache 2.0 - 435M params	Open weights	Not stated	Not stated
bge-reranker-baseBAAI - `BAAI/bge-reranker-base`	Open weightsMIT - 278M params	Open weights	Not stated	Not stated
Llama Nemotron Rerank 500m v2NVIDIA NIM - `nvidia/llama-nemotron-rerank-500m-v2`	Not stated	Self-hosted NIM container	8,192optimized NIM profiles	Not stated

Limits, languages and weights

Every current model. Context figures are scoped differently by each provider (per document, query plus document, or per record), so each one carries its scope. A document limit marked as a recommendation is not enforced by the API.

35 current reranker models with context, documents per request, languages, open weights and status
Model	Context (tokens)	Docs per request	Languages	Weights	Status
Rerank 4 ProCohere - `rerank-v4.0-pro`	32,000context length (models table lists 32k)	1,000recommended maximum, not a hard limit	Over 100 languages	API only inferred	GA
Rerank 4 FastCohere - `rerank-v4.0-fast`	32,000context length (models table lists 32k)	1,000recommended maximum, not a hard limit	Over 100 languages	API only inferred	GA
Rerank 3.5Cohere - `rerank-v3.5`	4,096per query plus document	1,000recommended maximum, not a hard limit	Not stated	API only inferred	GAinferred
Rerank English 3.0Cohere - `rerank-english-v3.0`	4,096per query plus document	1,000recommended maximum, not a hard limit	English	API only inferred	GAinferred
Rerank Multilingual 3.0Cohere - `rerank-multilingual-v3.0`	4,096per query plus document	1,000recommended maximum, not a hard limit	Same languages as embed-multilingual-v3.0 (non-English)	API only inferred	GAinferred
rerank-3Voyage AI - `rerank-3`	32,000query plus any single document	1,000hard limit	Not stated	API only inferred	GA
rerank-3-liteVoyage AI - `rerank-3-lite`	32,000query plus any single document	1,000hard limit	Not stated	API only inferred	GA
Amazon Rerank 1.0Amazon Bedrock - `amazon.rerank-v1:0`	Not stated	1,000hard limit (sources array of the Bedrock Rerank API)	Not stated	API only inferred	GAinferred
Cohere Rerank 3.5 (Bedrock)Amazon Bedrock - `cohere.rerank-v3-5:0`	4,000context window (model card says 4K)	1,000hard limit (sources array of the Bedrock Rerank API)	Not stated	API only inferred	GA
semantic-ranker-default-005Google Cloud - `semantic-ranker-default-005`	1,024per record (title plus content)	1,000hard limit	25 languages	API only inferred	GAinferred
semantic-ranker-fast-005Google Cloud - `semantic-ranker-fast-005`	1,024per record (title plus content)	1,000hard limit	25 languages	API only inferred	GAinferred
semantic-ranker-default-004Google Cloud - `semantic-ranker-default-004`	1,024per record (title plus content)	1,000hard limit	25 languages	API only inferred	GAinferred
semantic-ranker-fast-004Google Cloud - `semantic-ranker-fast-004`	1,024per record (title plus content)	1,000hard limit	25 languages	API only inferred	GAinferred
semantic-ranker-default-003Google Cloud - `semantic-ranker-default-003`	512per record (title plus content)	1,000hard limit	25 languages	API only inferred	GAinferred
semantic-ranker-default-002Google Cloud - `semantic-ranker-default-002`	512per record (title plus content)	1,000hard limit	English only	API only inferred	GAinferred
mxbai-rerank-v3.1-listwiseMixedbread - `mixedbread-ai/mxbai-rerank-v3.1-listwise`	Not stated	Not stated	Not stated	Not stated	GA
mxbai-rerank-large-v2Mixedbread - `mixedbread-ai/mxbai-rerank-large-v2`	Not stated	Not stated	100+ languages	Open weightsApache 2.0 - 1.54B params	GAinferred
mxbai-rerank-base-v2Mixedbread - `mixedbread-ai/mxbai-rerank-base-v2`	Not stated	Not stated	100+ languages	Open weightsApache 2.0 - 494M params	GAinferred
mxbai-rerank-large-v1Mixedbread - `mixedbread-ai/mxbai-rerank-large-v1`	Not stated	Not stated	English	Open weightsApache 2.0 - 435M params	GAinferred
mxbai-rerank-large-v2 (Together dedicated)Together AI - `mixedbread-ai/mxbai-rerank-large-v2`	8,000per document (API-level statement)	Not stated	Not stated	Open weightsApache 2.0 - 1.54B params	GAinferred
Llama-Rank-V1 (Together dedicated)Together AI - `Salesforce/Llama-Rank-V1`	8,000per document (API-level statement)	Not stated	Not stated	Not stated	GAinferred
qwen3-rerankAlibaba Cloud Model Studio - `qwen3-rerank`	4,000per query or document; 120,000 per request	500hard limit	Over 100 major languages	Not stated	GAinferred
qwen3-vl-rerankAlibaba Cloud Model Studio - `qwen3-vl-rerank`	8,000per item; 120,000 per request	100hard limit for text documents (40 images, 4 videos)	33 major languages	Not stated	GAinferred
gte-rerank-v2Alibaba Cloud Model Studio - `gte-rerank-v2`	4,000per item; 30,000 per request	500hard limit	Over 50 languages	Not stated	GAinferred
Llama Nemotron Rerank 1B v2NVIDIA NIM - `nvidia/llama-nemotron-rerank-1b-v2`	8,192query plus passage, optimized NIM profiles	Not stated	Evaluated on 26 languages	Open weightsOpenMDW-1.1 - 1.24B params	GA
Llama Nemotron Rerank VL 1B v2NVIDIA NIM - `nvidia/llama-nemotron-rerank-vl-1b-v2`	8,192optimized NIM profiles	Not stated	Multilingual (no count stated)	Open weightsNVIDIA Open Model License - 1.68B params	GA
Llama Nemotron Rerank 500m v2NVIDIA NIM - `nvidia/llama-nemotron-rerank-500m-v2`	8,192optimized NIM profiles	Not stated	Not stated	Not stated	GAinferred
bge-reranker-v2-m3BAAI - `BAAI/bge-reranker-v2-m3`	Not stated	Not stated	Multilingual (no count stated)	Open weightsApache 2.0 - 568M params	GAinferred
bge-reranker-v2-gemmaBAAI - `BAAI/bge-reranker-v2-gemma`	Not stated	Not stated	Multilingual (no count stated)	Open weightsApache 2.0 - 2.51B params	GAinferred
bge-reranker-v2-minicpm-layerwiseBAAI - `BAAI/bge-reranker-v2-minicpm-layerwise`	Not stated	Not stated	Multilingual (no count stated)	Open weightsApache 2.0 - 2.72B params	GAinferred
bge-reranker-largeBAAI - `BAAI/bge-reranker-large`	Not stated	Not stated	Chinese and English	Open weightsMIT - 560M params	GAinferred
bge-reranker-baseBAAI - `BAAI/bge-reranker-base`	Not stated	Not stated	Chinese and English	Open weightsMIT - 278M params	GAinferred
Qwen3-Reranker-0.6BQwen - `Qwen/Qwen3-Reranker-0.6B`	32,000context length (card says 32k)	Not stated	100+ languages	Open weightsApache 2.0 - 596M params	GAinferred
Qwen3-Reranker-4BQwen - `Qwen/Qwen3-Reranker-4B`	32,000context length (card says 32k)	Not stated	100+ languages	Open weightsApache 2.0 - 4.02B params	GAinferred
Qwen3-Reranker-8BQwen - `Qwen/Qwen3-Reranker-8B`	32,000context length (card says 32k)	Not stated	100+ languages	Open weightsApache 2.0 - 8.19B params	GAinferred

Older models

Earlier generations their providers still serve or alias to a newer model. Prices stay in the provider's unit and are not ranked against the current tables above.

7 older reranker models still served, with price in the provider's unit, context and status evidence
Model	Price (provider's unit)	Context (tokens)	Query limit (tokens)	Status, from the source
rerank-2.5Voyage AI - `rerank-2.5`	$0.05 per 1M tokensdocs.voyageai.com	32,000query plus any single document	8,000	“The following are our earlier models, which are still accessible from our API.”
rerank-2.5-liteVoyage AI - `rerank-2.5-lite`	$0.02 per 1M tokensdocs.voyageai.com	32,000query plus any single document	8,000	“The following are our earlier models, which are still accessible from our API.”
rerank-2Voyage AI - `rerank-2`	$0.05 per 1M tokensdocs.voyageai.com	16,000query plus any single document	4,000	“The following are our earlier models, which are still accessible from our API.”
rerank-2-liteVoyage AI - `rerank-2-lite`	$0.02 per 1M tokensdocs.voyageai.com	8,000query plus any single document	2,000	“The following are our earlier models, which are still accessible from our API.”
rerank-1Voyage AI - `rerank-1`	$0.05 per 1M tokensdocs.voyageai.com	8,000query plus any single document	2,000	“The following are our earlier models, which are still accessible from our API.”
rerank-lite-1Voyage AI - `rerank-lite-1`	$0.02 per 1M tokensdocs.voyageai.com	4,000query plus any single document	1,000	“The following are our earlier models, which are still accessible from our API.”
mxbai-rerank-v3-listwiseMixedbread - `mixedbread-ai/mxbai-rerank-v3-listwise`	Not stated	Not stated	Not stated	“No longer supported. Requests using this name are served by mixedbread-ai/mxbai-rerank-v3.1-listwise; use that model instead.”

Voyage AI

Per-request total token cap: "cannot exceed 600K for rerank-3, rerank-3-lite, rerank-2.5, rerank-2.5-lite, rerank-2 and rerank-2-lite; and 300K for rerank-1 and rerank-lite-1" (docs). The Batch API is listed at a 33% discount compared to standard endpoints. Docs disagree on free tokens: the Rerankers section says "The first 200 million tokens for rerank-2.5, rerank-2.5-lite, rerank-2, and rerank-2-lite are free for each account", while the Older models table, which lists those same models, says "we do not offer free tokens for them." The current-model table lists 200 million free tokens for rerank-3 and rerank-3-lite.

Mixedbread

Mixedbread no longer documents a standalone rerank endpoint; reranking is a search option on its Stores API, and rerank pricing is an add-on per 1K Store queries.

How each provider bills

Each provider's billing unit, quoted where it publishes a definition, and the caveats we recorded. Every provider page lists each model with its verbatim quotes, sources and notes.

Cohere
Billing unit: search unit
“A single search unit is defined as one query with up to 100 documents to be ranked.”
cohere.com
Rerank documents are chunked internally; the models page says Rerank "will automatically chunk documents longer than 510 tokens" while the context column lists 4k or 32k. Billing unit quote: "A single search unit is defined as one query with up to 100 documents to be ranked." (cohere.com/pricing FAQ). The rerank-v2.0 family was shut down 2025-04-30 per the deprecations page.
Voyage AI
Per-request total token cap: "cannot exceed 600K for rerank-3, rerank-3-lite, rerank-2.5, rerank-2.5-lite, rerank-2 and rerank-2-lite; and 300K for rerank-1 and rerank-lite-1" (docs). The Batch API is listed at a 33% discount compared to standard endpoints. Docs disagree on free tokens: the Rerankers section says "The first 200 million tokens for rerank-2.5, rerank-2.5-lite, rerank-2, and rerank-2-lite are free for each account", while the Older models table, which lists those same models, says "we do not offer free tokens for them." The current-model table lists 200 million free tokens for rerank-3 and rerank-3-lite.
Amazon Bedrock
Billing unit: query (up to 100 document chunks)
“A query is a single call to the reranker model that can contain up to 100 document chunks.”
docs.aws.amazon.com
Rerank billing unit (docs): "A query is a single call to the reranker model that can contain up to 100 document chunks." The Bedrock pricing footnote (checked 2026-10-18) says: "each document can only contain upto 512 tokens (inclusive of the query and document’s total tokens), and if the token length is higher than 512 tokens, it is broken down into multiple documents." It also says a request with 350 documents "will be treated as 4 queries."
Google Cloud (Vertex AI ranking API)
Status GA is inferred: the docs list every model as available with discontinuation "To be determined" and do not label any as preview. Records that exceed the per-record token limit are truncated, not rejected.
Mixedbread
Mixedbread no longer documents a standalone rerank endpoint; reranking is a search option on its Stores API, and rerank pricing is an add-on per 1K Store queries.
Together AI
Rerank is dedicated-endpoint only on 2026-10-18; there is no per-token rerank price.
Alibaba Cloud Model Studio
Region matters: qwen3-rerank appears under the Singapore tab, qwen3-vl-rerank and gte-rerank-v2 under the Beijing tab (tab mapping inferred from page order).
NVIDIA NIM
NIM containers are self-hosted; the support matrix lists no per-call price, so pricing is null.
BAAI (open weights)
Languages come from the model list on the bge-reranker-v2-m3 card. Status GA means released on Hugging Face (inferred).
Qwen (open weights)
32k is the card's rounded figure; it is recorded here as 32000. Alibaba's hosted qwen3-rerank API is listed separately under Alibaba Cloud Model Studio.

Method

How to read these numbers

Read from official provider docs, pricing pages and Hugging Face model cards on October 18, 2026, with every quote string-matched against the page that day. Prices and limits change; check the linked source before you commit.

Not a quality benchmark: Nothing here measures ranking quality or latency. Every value is a first-party fact: price, context, document limits, languages and license. Shortlist on these facts, then evaluate on your own queries and documents.
Prices in the provider's own unit: Per 1M tokens, per 1,000 queries, a per-query add-on and hourly dedicated capacity are different units. Each gets its own table and rows are sorted only inside one table; no price is converted.
Inferred means implied, not labelled: An inferred status is one the source implies (listed as available, no GA label). An inferred "API only" means no official open weights were found in the sources checked for an API model.
Not stated means not stated: When the checked source does not give a value, the cell says so instead of guessing. Cohere's pay-as-you-go price per search unit is one: it was not in the page text we read.

FAQ

What is the cheapest reranker API?

There is no single answer, because providers bill in different units that do not convert into each other. Priced per 1M tokens, the lowest current rate is $0.02 per 1M tokens for rerank-3-lite (Voyage AI). Priced per 1,000 queries, the lowest current rate is $1.00 per 1,000 queries, for Amazon Rerank 1.0 (Amazon Bedrock), semantic-ranker-default-002 (Google Cloud), semantic-ranker-default-003 (Google Cloud) and 4 more. Cohere's pay-as-you-go price per search unit is not stated in the page we read, and Together AI runs rerank models only on dedicated endpoints. A low price says nothing about ranking quality, so test a shortlist on your own queries.

How does Cohere bill Rerank?

In search units: "A single search unit is defined as one query with up to 100 documents to be ranked." The pay-as-you-go price per search sits in a pricing-page tab that was not present in the page text we read, so this directory shows it as Not stated. Dedicated Model Vault instances are listed per instance per hour: Rerank 4 Pro Medium $5.00, Rerank 4 Pro Large $10.00, Rerank 4 Fast Medium $5.00 and Rerank 3.5 Medium $5.00.

What counts as a query on Amazon Bedrock and Google Cloud?

Amazon Bedrock: "A query is a single call to the reranker model that can contain up to 100 document chunks." Its pricing page adds that a document over 512 tokens, query included, is split into several documents. Google Cloud's ranking API: Pay-as-you-go column; the next two columns ($0.90 and $0.80) are 1-year and 3-year savings-plan rates. A query is up to 100 documents; every further 100 documents counts as another query.

How are Voyage AI rerankers billed?

Per 1M tokens. Billed on total processed tokens: (query tokens x number of documents) + tokens in all documents. Voyage's own docs disagree on free tokens for its older rerankers: the Rerankers section offers the first 200 million tokens free for rerank-2.5, rerank-2.5-lite, rerank-2 and rerank-2-lite, while the Older models table says no free tokens are offered for them. This directory shows both rather than picking one.

Which rerankers can I run myself?

13 distinct checkpoints publish open weights: Qwen3-Reranker-8B, Qwen3-Reranker-4B, bge-reranker-v2-minicpm-layerwise, bge-reranker-v2-gemma, Llama Nemotron Rerank VL 1B v2, mxbai-rerank-large-v2, Llama Nemotron Rerank 1B v2, Qwen3-Reranker-0.6B and 5 more. Licenses: Apache 2.0, NVIDIA Open Model License, OpenMDW-1.1 and MIT. Some open-weight models carry extra terms from their base model, noted on each provider page.

How many documents can I send in one rerank request?

13 current models state a hard limit, from 100 to 1,000 documents. Cohere's 1,000 is a recommended maximum, not a hard limit. Billing units are smaller than these limits on several APIs: Bedrock and Google Cloud count every 100 documents as another query.

Is this a reranker benchmark?

No. Every value is a first-party fact copied from provider docs, pricing pages and Hugging Face cards: prices, context limits, document limits, languages and licenses. Nothing here measures ranking quality or latency.

What do inferred and Not stated mean?

Not stated means the checked source did not give the value, so it is left blank rather than guessed. Inferred marks a status the source implies but does not label (a model listed as available with no GA badge), or an API model with no open weights found in the sources checked.

Last verified 2026-10-18 from each provider's own docs, pricing pages and Hugging Face cards; every quote was string-matched against the page that day. Prices are shown in the provider's own unit and change without notice, so check the linked source before you commit a budget.

Snapshot of October 18, 2026: 42 models from 10 providers, 21 with a per-call price, 13 distinct open-weight checkpoints, 7 older.

Newsletter

One email per week. New model releases, retrieval and pricing updates and what they mean for what you build.

Reranker models compared

Pricing by billing unit

Per 1M tokens

Per 1,000 queries

Add-on per 1K Store queries

Dedicated hourly only (Cohere Model Vault)

Dedicated endpoint only (Together AI)

Self-host (open weights or NIM containers)

Limits, languages and weights

Older models

Voyage AI

Mixedbread

How each provider bills

Cohere

Voyage AI

Amazon Bedrock

Google Cloud (Vertex AI ranking API)

Mixedbread

Together AI

Alibaba Cloud Model Studio

NVIDIA NIM

BAAI (open weights)

Qwen (open weights)

How to read these numbers

FAQ

What is the cheapest reranker API?

How does Cohere bill Rerank?

What counts as a query on Amazon Bedrock and Google Cloud?

How are Voyage AI rerankers billed?

Which rerankers can I run myself?

How many documents can I send in one rerank request?

Is this a reranker benchmark?

What do inferred and Not stated mean?

Embedding models compared

What is RAG?

Vector databases compared

Local models

Pricing by billing unit

Per 1M tokens

Per 1,000 queries

Add-on per 1K Store queries

Dedicated hourly only (Cohere Model Vault)

Dedicated endpoint only (Together AI)

Self-host (open weights or NIM containers)

Limits, languages and weights

Older models

How each provider bills

How to read these numbers

FAQ

What is the cheapest reranker API?

How does Cohere bill Rerank?

What counts as a query on Amazon Bedrock and Google Cloud?

How are Voyage AI rerankers billed?

Which rerankers can I run myself?

How many documents can I send in one rerank request?

Is this a reranker benchmark?

What do inferred and Not stated mean?

Know when a reranker API changes

Keep exploring

Embedding models compared

What is RAG?

Vector databases compared

Local models