Reranker API pricing, unit by unit, with the latency nobody prints

Reranker APIs do not share a pricing model. Some bill by the search. Some bill by the token. Some are free up to a rate limit, and some rent you an instance by the hour. So two rate cards can differ by a unit rule alone, and a page of prices without units cannot answer your question. This page fixes that. It re-read every vendor's own pricing surface on 28 October 2026. It pulls the billable unit rule from each. It states every competitor rate as a dated relation to a flat per-search rate, never as a number. And it gives you latency figures with their dates. This platform serves one reranker, bge-reranker-v2-m3, at a flat rule: one request is one metered search, whatever the document count.

Every number below names its owner and its date. The Cohere facts come from its pricing page, its pricing docs, and the Rerank 5 post, read 28 October 2026. The Voyage facts come from its pricing docs, the same day. The cloud rows come from AWS Bedrock pricing and Google Cloud Agent Search pricing, the same day. Marketplace rows are labeled database-reported and dated. Where a rate could not be read on a public page, this page prints a dash instead of a guess. The pages ranking today list rates without unit rules. They promise latency comparisons while printing no latency numbers. This page does both jobs.

Last verified: 2026-10-28

Why reranker prices do not line up

Ask four vendors what one rerank call costs and you get four different questions back. What is a call. What is a document. What happens when a document is long. The pricing model decides all three. The market today runs four models at once.

The first model bills by the search. You send a query plus a shortlist. You get ranked results. The bill counts one search. AWS Bedrock prices Cohere Rerank 3.5 this way, and so does this platform. The second model bills by the token. Voyage AI counts query tokens times document count, plus all document tokens, and prices that total per million. Cohere's newest generation, Rerank 5, moved to per-token rates on its launch post. Its standing docs still describe search-based pricing. The market itself is mid-shift. The third model is the free tier. NVIDIA's build surface serves rerank models free to prototype under rate limits, with no per-token billing. The fourth model rents capacity. Cohere's Model Vault sells rerank instances by the hour. Together AI offers rerank models only through dedicated endpoints.

None of these four is dishonest. But they cannot be compared by reading rate cards side by side, because the unit definitions differ under the rates. A vendor that bills 100 documents as four queries is not cheaper than one that bills them as one search. The headline rates cannot tell you that. The unit table below is the comparison. Read it before any rate.

What each provider bills: the unit table

Each row below is the provider's own unit rule, read from its own surface on the date shown. This is the table no ranking page prints. It is the one that makes the rates comparable.

Three rows deserve a second look. The AWS row means a 350-document request bills as four queries, not one. The Google row climbs the same ladder, one count per 100 documents. The Cohere row splits long documents into billable chunks at a 500-token line. This platform's row stays flat. One request is one search, at any document count. The corpus size is recorded but never priced. Do your shortlists run long? Do your documents run past a few hundred tokens? Then these unit rules move your bill more than any headline rate.

The rates, as relations

This page states competitor rates only as directions against a flat per-search rate, each named and dated. It never prints a competitor's price figure. A figure without its unit rule misleads. A dated relation carries the real information. This platform's own rate renders below from the platform's model data, so it is current whenever you read it.

Budget math on named variables, for the flat lane:

monthly cost here  =  ( Q / 1,000 ) x R

Q:  rerank searches in the month; one request is one search, any document count
R:  per-1k-search rate from the fenced row above

And the translator for a per-token rival, using its own unit rule:

a per-token rival's request cost  =  ( T / 1,000,000 ) x R_t

T:   tokens the rival counts for the request, by its unit rule
R_t: its per-million-token rate, from its own card

The two formulas meet at one flip point. That point is the only honest way to compare the models:

flip point  =  1,000,000 x ( R / 1,000 ) / R_t   tokens per request

Below the flip point the per-token vendor bills less than one search here.
Above it, the flat search wins. No vendor card prints this number.

One sensitivity note. T scales with document length and shortlist size. Many short documents sit far below the flip point. Few long documents can sit far above it. Run the formula with your real shortlist shape before believing either side of a pricing argument.

The relations, each named and dated:

• Voyage AI, per its pricing docs read 28 October 2026: at its own published estimate shape of 100 documents at 500 tokens each, the rerank-3 rate lands above one search on this platform, and the rerank-3-lite rate lands below it. Same shape, same vendor, both directions exist inside one rate card.
• Cohere, per the Rerank 5 post and pricing docs read 28 October 2026: at Cohere's own benchmark shape of 96 documents at 1,000 tokens each, the Rerank 5 rates land above one search on this platform. On short shortlists of small documents, they land below. The direction flips inside Cohere's own card, which is exactly why a single verdict would mislead.
• Cohere's pay-as-you-go rate card: dash. The card sits behind a consent wall that no static read reached on 28 October 2026, so this page prints no rate and no direction for it. The search-unit definition above is from the public FAQ on the same page.
• Cohere's older generations, per a third-party marketplace collection read 28 October 2026, database-reported: the per-search listing for Rerank 4 Pro sits above this platform's per-search rate; the listing for the older v3.5 sits below it.
• AWS Bedrock, per its pricing page read 28 October 2026: the per-1k-query rate for Cohere Rerank 3.5 sits above this platform's per-1k-search rate at the flat shape. Its unit rule bills a 350-document request as four queries. This platform bills one search.
• Google Cloud, per its pricing page read 28 October 2026: the Ranking API's per-1k-count rate sits below this platform's per-1k-search rate at the flat shape. It climbs one count per 100 documents.
• Amazon-native rerank, per the LiteLLM catalog row checked 27 October 2026, database-reported: the per-1k-query rate sits below this platform's flat rate at the same unit.
• Novita AI, per its catalog row checked 27 October 2026, database-reported, and its provider post: the same bge-reranker-v2-m3 model this platform serves, priced per token instead. At the 100-document, 500-token shape it lands below one search here. Past the flip point it lands above. Same model, two billing models, both directions real.
• Qwen3 Reranker 8B, per the same marketplace collection read 28 October 2026, database-reported: the per-token listing sits above one search on this platform at that same 100 by 500 shape.
• Azure AI Foundry: dash, as noted in the unit table. Jina AI sells token packages rather than a per-request rate, so no direction applies. Together AI offers dedicated endpoints only, and its docs read 28 October 2026 carry no per-request rate to relate.

Notice the pattern across the list. At a named shape, most cards land on both sides of this platform's flat rate depending on the shape. That is not a marketing claim. It is arithmetic on unit rules. It is the reason this page prints relations at shapes instead of a verdict.

Latency: the numbers you can plan around

Rerank latency has a physical cause. A reranker is a cross-encoder: it scores the query against every document on your shortlist, one pair at a time. More documents means more scoring. Longer documents means more scoring per pair. So any latency number is a number for a specific shortlist shape. A bare figure with no shape is close to useless.

If the question under the pricing one is whether a cross-encoder beats embedding retrieval alone, the cross-encoder vs bi-encoder guide works that tradeoff end to end.

The pages ranking today mostly offer no figure at all. The marketplace collection that ranks today opens by promising to compare pricing, latency, and quality. The page read 28 October 2026 prints zero latency numbers. Voyage's pricing docs print none. AWS and Google print none. Jina's page answers the latency question with a note that it depends on the input size, which is true and unhelpful. A pricing calculator site carries a qualitative Low label for Cohere Rerank 5. That label is the closest thing to a latency figure anywhere in the results. It is a word, not a number.

What exists, with dates:

Read the Cohere numbers as a vendor measuring its own model on its own hardware: useful scale, single source. Read this platform's numbers as planning inputs from the 9 September 2026 window, not as promises. The window's p95 of about 0.9 seconds single flight, at an 8-in-flight ceiling, is a number you can size a retrieval stage around. That is more than most rate cards give you.

One cost note ties back to the unit table. On this lane, latency grows with your shortlist but the bill does not. One request is one search at any document count. On a per-token lane, the same growth hits both your latency and your bill. Long shortlists are a latency question everywhere. They are only a cost question where the unit rule says so.

Free tiers, read closely

Free rerank exists, and the boundaries are the product.

NVIDIA's build surface is the widest free door. All models are free to prototype. Most models run at up to 40 requests per minute. There is no per-token billing. The NVIDIA API trial terms apply, per its model page read 28 October 2026. For a prototype that is real capacity. For production it is a rate limit you will meet. The two NVIDIA rerank models also list at zero on a third-party marketplace collection, database-reported, read the same day.

Voyage grants the first 200 million tokens free on the rerank-2.5 family, per its pricing docs read 28 October 2026. That is a generous one-time grant, and it is tokens, so the burn rate follows your shortlist shape. Jina's free key runs at 100 requests per minute and 100 thousand tokens per minute, per its reranker page read the same day. Paid tiers raise that to 500 requests and 2 million tokens. A premium tier reaches 5,000 and 50 million.

Two capacity options sit at the other end. Cohere's Model Vault rents rerank instances by the hour, a fixed cost for a fixed throughput target. Together AI offers rerank models through dedicated endpoints only, per its docs read 28 October 2026. Inside AWS, the Knowledge Bases managed re-ranker is included with the service, per the Bedrock pricing page read the same day. Free here means bundled. It does not mean the service costs nothing.

The honest summary: free tiers answer the question "does rerank help my pipeline", and paid lanes answer "what does rerank cost at my volume". The second question is the one the unit table and the relations above answer.

Does OpenAI have a rerank API?

No, and the searches asking for one deserve a straight answer. OpenAI's platform reference, read 28 October 2026, lists its endpoint families: chat, completions, embeddings, responses, realtime, and more. A search across the rendered reference finds zero occurrences of rerank. The platform's llms.txt listing returns not found. Its sitemap stub carries nothing. The search results for "rerank api openai" surface no OpenAI page at all. An adapter library ranks first. A mirror of its docs ranks second. Other vendors follow.

That absence is why the query exists. Engineers assume a rerank endpoint belongs next to the embeddings endpoint, search for it, and find adapters. The adapters are real software. But the endpoint they wrap is not OpenAI's, and the bill does not arrive on an OpenAI invoice.

If you wanted rerank on an OpenAI-shaped contract, look at the lanes on this page instead. A third-party hosted reranker, such as this platform's. A cloud marketplace row, such as AWS Bedrock's. Or self-hosting an open reranker. The rerank lane page maps the first option. Its model is open, Apache 2.0, so the self-host exit stays open too.

Pricing a rerank lane on this platform

As of 28 October 2026, this platform serves one reranker, bge-reranker-v2-m3, on its rerank route. The lane is live and the metering rule is flat. One request is one metered search, whatever the document count. The corpus size is recorded but never priced. There is no token arithmetic to run. There is no chunk ladder to climb.

The model itself is a 568M-parameter cross-encoder on an XLM-RoBERTa base, released March 2024 under Apache 2.0, with an 8,192-token window per query-document pair. Scores are sigmoid values between 0 and 1. BAAI's own family table describes it as a lightweight cross-encoder. It has strong multilingual coverage, it is easy to deploy, and it runs fast inference. The community agrees at scale. The model host's counter read 28 October 2026 shows about 14.7 million downloads in the last month.

The request contract is strict, four fields, on one route. Send a model, a query, a documents list, and top_n. The parser rejects anything else, which catches integration bugs before they bill. A first call looks like this:

curl https://api.ironstratum.com/v1/rerank \
  -H "Authorization: Bearer $KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bge-reranker-v2-m3",
    "query": "what released the four-wheeled demo robot",
    "documents": [
      "the demo robot ran a vision-language-action policy on an onboard computer",
      "the warehouse arm used a cross-encoder reranker before retrieval"
    ],
    "top_n": 2
  }'

Results arrive as an ordered list of indexes with relevance scores, highest first. Retries are safe: send an Idempotency-Key header and a replayed request bills once, per the idempotency docs. The wallet is prepaid. A refused request bills nothing. That hard stop prevents the surprise-invoice failure mode the bill shock guide covers in depth. The model page carries the full contract and the benchmark pins. The rerank lane page maps the route. The models index is the live catalog. The pricing page always carries the current rate. When you are ready, sign up and make the call above with your own key.

What to check before you commit

Work the list in order and stop at the first red flag.

1. The unit rule at your shortlist shape. Count your documents, then apply each vendor's rule from the table above. A 350-document request is one search here, four queries on AWS, four counts on Google. Price that difference before anything else.
2. The chunk rule on long documents. If your documents run past 500 tokens, Cohere's rule bills each chunk as a document. If they run past 512 on AWS, the same thing happens. On a per-token lane, length bills by arithmetic instead.
3. Latency at your document length. Scale from the dated figures above by pair count, or send 20 timed requests to the lane. Rerank latency follows the shortlist; test yours, not a benchmark's.
4. The free-tier boundary if you are prototyping. 40 requests per minute, 200 million tokens once, 100 requests per minute: these are three different ceilings, and crossing one silently is how prototypes become incidents.
5. Retry behavior. Confirm idempotency before you wire retries into a loop, or a timeout storm becomes a bill. This lane's Idempotency-Key behavior is on the idempotency docs.

Common questions

Is rerank billed per search or per token?

Both exist, and that split is the market's current state. This platform, AWS Bedrock, and Azure bill per search. Voyage, Novita, Jina, and Cohere's Rerank 5 generation bill per token. The unit table above carries each provider's exact rule, read 28 October 2026. The flip-point formula in the relations section converts between them at your shortlist shape.

What counts as one search?

Here: one request, any number of documents, corpus size recorded unpriced. At Cohere: one query with up to 100 documents, with documents over 500 tokens splitting into billable chunks. At AWS: up to 100 document chunks per query, with 350 documents billing as four queries. At Google: up to 100 documents per count. Four providers, four different answers, all from their own pages on the same day.

Why do per-token rates punish long documents?

Because the token count grows with document length, and the rate multiplies the count. A 100-document shortlist at 500 tokens each bills about 50,000 tokens on a per-token lane, whatever the query. The same shortlist at 5,000 tokens each bills about ten times that. On a per-search lane like this one, both bills are identical: one search. Length is a latency question everywhere. It is a cost question only where tokens are the unit.

Is there a free rerank API?

Yes, within limits. NVIDIA's build surface serves rerank models free to prototype at up to 40 requests per minute on most models, under trial terms. Voyage grants 200 million tokens once on the rerank-2.5 family. Jina's free key runs 100 requests per minute and 100 thousand tokens per minute. All three, read 28 October 2026. For production volumes, every lane on this page bills.

Does OpenAI offer a rerank endpoint?

No. Its platform reference, read 28 October 2026, lists no rerank endpoint, and its own surfaces carry no such product. Search results for the query sell adapters and third parties. The lanes that actually serve hosted rerank are covered in the unit table, and this platform's is one of them.

What latency should I plan for a rerank stage?

Plan from dated figures at a named shape. On this lane, from the 9 September 2026 window: first response p95 about 0.9 seconds single flight, 8 concurrent requests in flight. For Cohere's Fast tier, its vendor benchmark reports about 187 documents per second on one A100 at 1,000-token documents. Both are planning inputs, not promises. Scale by your pair count and your token length, then measure once on your own shortlist.

The short version

Reranker pricing cannot be compared by rate cards. The billable units differ: a search, a token, a free-tier limit, an hourly instance. Pull the unit rule first, then the rate. This page re-read every vendor's own surface on 28 October 2026. It printed the rules in one table, the rates as dated relations, and the latency figures that exist, with their dates. This platform's lane is the flat corner of that market. bge-reranker-v2-m3, one request is one search at any document count. Prepaid wallet. Idempotent retries. The live rate sits on the pricing page. The route is mapped on the rerank lane page. If rerank belongs in your retrieval stage, sign up and price it on your real shortlist, not on a headline.

Last verified: 2026-10-28