Cohere released Rerank 5 on Oct. 21, 2026, adding passage-level selection to its enterprise search models. The change gives teams building AI search and agent systems another way to control which information reaches a language model and contributes to its processing bill.
Rapid News Brief — The release includes Pro and Fast versions. Cohere positions Pro for more demanding retrieval tasks and Fast for workloads where response time and throughput carry greater weight. Those product descriptions do not establish which version will perform best on a customer’s own documents.
Selecting passages, not only documents
The new Context Focus option identifies relevant spans within a document, rather than stopping at a document-level ranking. Cohere says developers can pass those selected passages to a language model or highlight them for users. The feature is off by default, according to the release notes.
The company’s launch announcement lists API pricing of 5 cents per million tokens for Pro and 2 cents for Fast. It says the models are generally available through Cohere’s API, Model Vault, Microsoft Foundry and Amazon SageMaker, as well as to North users.
The separate developer release notes say an upgrade does not require re-indexing or re-embedding an existing document collection because reranking operates on raw text at query time. That narrows one part of the migration effort; it does not remove the need to test the application’s results.
The business test is complete answer quality
For enterprise teams, selecting less text is useful only if the application preserves the evidence needed to answer a question. A shorter input can omit an important exception or definition as well as remove irrelevant material. A practical evaluation should therefore compare completed answers, supporting passages and total processing effort on representative questions.
That is a testing framework, not an independent assessment of Rerank 5. Cohere reports performance improvements in its own evaluations, but a launch benchmark does not establish savings or accuracy for a particular company. Buyers should distinguish the listed price of the retrieval model from the full cost and quality of the workflow using it.