Futureweb

Artificial Intelligence

Cohere's Rerank 5 scores Cohere's own chart. Context Focus is optional.

Cohere's October 21 post introduces Rerank 5 in Pro and Fast sizes. The benchmark table is Cohere's. Context Focus returns scored spans in the same call, and it is optional.

The short version

Cohere's Rerank 5 is two models, Pro and Fast, generally available on Cohere's API and named partner platforms. The scores Cohere publishes are Cohere's, using Cohere's own metric. Context Focus is an optional span list in the same call. It is not a second lab's ranking.

Two model names, one company

Cohere's October 21 blog introduces Rerank 5 as two models. Pro is rerank-v5.0-pro. Fast is rerank-v5.0-fast. The docs page, which is the same company, describes Pro at about 2 billion parameters and Fast at about 500 million, a 32,768 token context that counts the query and the document together, and support for more than 100 languages. Scores run from 0 to 1. The changelog and the blog are one vendor's documents. MarkTechPost is the other account on the wire.Exa 3 ranks its own tests.

Where Cohere says it is generally available

The blog says Rerank 5 is generally available on the Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker, and North. A platform name on a launch post is a distribution claim. It is not a customer's case study, and it is not a statement that every tenant on those platforms has the model turned on. A reader who needs a region or a quota reads Cohere's current docs, which can move after this page. Asked directly, Does a score between 0 and 1 decide a cutoff? Cohere says the models score from 0 to 1. A cutoff is the caller's choice. This page does not print one.

The price is the blog's price

Cohere's blog lists $0.05 per million tokens for Pro and $0.02 per million tokens for Fast. That is a price on a launch post. It is not a quote this page negotiated, and it is not a promise that the figure is still on the billing page next month. The two rates belong to the two model names. Mixing them into one average would invent a third product. Asked directly, Whose number is the BRIGHT score? Cohere's. The company says Pro scored 82.2 on BRIGHT under its RCP-nDCG@10. That is not an independent lab.

The chart is Cohere's chart

Cohere says it evaluated the models on ViDoRe V3, nanoMTEB, BRIGHT, and BEIR, and that Pro scored 82.2 on BRIGHT and placed first on 39 of 46 datasets. The metric it names is Cohere's RCP-nDCG@10. A company scoring its own release on its own metric is a launch table. It is not a result a second lab signed. This page will not treat first place on that table as an independent ranking. MarkTechPost does not supply a second table.what a training cutoff does not mean.

Context Focus is optional

Cohere says Context Focus returns scored character spans in the same call, and that a caller can leave it off. A span list is an extra field on a score the model already returns. It does not change who wrote the benchmark. This page will not paste the request body, a threshold, or a sample span. Those are integration details. The fact that belongs here is that the feature is optional and that the score beside it is still Cohere's.

What a rerank call does not settle

A rerank score orders documents the caller already retrieved. It does not decide which corpus was searched, who is allowed to see a document, or whether the top span is true. Chunking still happens before retrieval. A higher score on Cohere's chart does not move a document into a corpus it was never in. People who already have a candidate list are the users the product is aimed at. Asked directly, What price does the blog list? Cohere's blog lists $0.05 per million tokens for Pro and $0.02 for Fast. A later price page would replace that sentence.chunk before retrieval.

Who is in scope

Developers comparing a Cohere rerank model with a score they can reproduce themselves. Procurement readers who need to know the table is the vendor's. People who have not retrieved any documents yet are outside the product's job. No customer win is in the sources. No independent leaderboard is in the sources either. Asked directly, Are the Rerank 5 blog and the changelog two companies? No. Both are Cohere. MarkTechPost is the other account, and it retells the post. The document named in the sources is the record this page will keep.

What stays on the record

October 21, 2026 Cohere blog. Pro and Fast. Docs: about 2 billion and about 500 million parameters, 32,768 token context, 100-plus languages, scores from 0 to 1. Blog prices: $0.05 and $0.02 per million tokens. GA on the surfaces the blog names. Context Focus optional. Benchmarks are Cohere's, including Pro at 82.2 on BRIGHT under Cohere's metric. MarkTechPost retells the post. Asked directly, What does Cohere say Context Focus returns? Cohere says Context Focus can return scored character spans in the same call. The feature is optional. This page does not paste a request body.what on-device inference keeps.

What would move the row

A second lab running the same sets under a metric it publishes, or a Cohere price page that replaces the blog's rates, would be a later document. A new model name would be a new row. Until then the record is Cohere's post, Cohere's docs, and one retelling. The short answer this page will stand on is this: Cohere's Rerank 5 is two models, Pro and Fast, generally available on Cohere's API and named partner platforms. The scores Cohere publishes are Cohere's, using Cohere's own metric. Context Focus is an optional span list in the same call. It is not a second lab's ranking.an evaluation a second lab can run.

Sources

The reports this brief is filing. Futureweb did not republish them.

  1. Cohere
  2. Cohere changelog
  3. MarkTechPost

Questions

Are the Rerank 5 blog and the changelog two companies?

No. Both are Cohere. MarkTechPost is the other account, and it retells the post.

What does Cohere say Context Focus returns?

Cohere says Context Focus can return scored character spans in the same call. The feature is optional. This page does not paste a request body.

Whose number is the BRIGHT score?

Cohere's. The company says Pro scored 82.2 on BRIGHT under its RCP-nDCG@10. That is not an independent lab.

What price does the blog list?

Cohere's blog lists $0.05 per million tokens for Pro and $0.02 for Fast. A later price page would replace that sentence.

Does a score between 0 and 1 decide a cutoff?

Cohere says the models score from 0 to 1. A cutoff is the caller's choice. This page does not print one.