Cohere Rerank 5 adds passage selection for AI search

Cohere announced Rerank 5 on October 21, 2026, in Pro and Fast versions. Its new Context Focus feature works inside documents: instead of forwarding every retrieved page, an application can select passages relevant to the question. That gives developers a way to balance shorter model inputs against retaining the evidence needed for an answer.

The model guide says passage selection is off by default. Developers enable relevant_spans; returned passages carry character positions and scores. Raising the threshold removes more text, while max_spans caps passages per document. If none qualifies, that result omits the passage field. This is intended for targeted questions: broad summarization needs the full document, and tables, forms and code divide less cleanly than prose.

The API reference describes an authenticated request containing a model identifier, a query and candidate text strings. The service returns ranked results with relevance scores. Two controls affect different stages: top_n limits how many results come back, whereas max_tokens_per_doc limits each input document. Its default is 4,096 tokens, and longer documents are truncated. Omitting top_n returns all ranked results; it does not remove the input truncation limit.

MarkTechPost’s report, published the following day, also covers the release. It explicitly identifies the benchmark figures as Cohere’s own measurements, without independent verification.