Listed in the changelog first

"Rerank 5," a new model from Canada's Cohere, which makes AI for businesses, appeared first in the company's developer changelog on Oct. 20 (Oct. 21 Japan time). @aitrackerbot, an X account that tracks the appearance of models, reported it. The company officially released it on Oct. 22.

@aitrackerbot · October 21, 2026

Rerank is a component used in systems that search internal documents and have generative AI answer questions (RAG). It reorders documents gathered by a search according to how relevant they are to the question and passes only the top ones to the generative AI.

High-accuracy Pro, fast Fast

Rerank 5 comes in two types, the high-accuracy "Pro" and the fast "Fast." Pro is suited to questions that require reasoning, long documents and documents in specialized fields, and the company says gains are largest in finance, medicine and scientific research. The company says Fast outperforms "Rerank 4 Pro," the higher-end version of the previous generation. The context window is 32,000 tokens, and it supports more than 100 languages. It can also handle structured data such as tables and JSON.

According to a South Korean blog, Pro scored 82.2 on BRIGHT, a benchmark for searches that require reasoning, topping Rerank 4 Pro's 69.8. Pricing per 1 million tokens is $0.05 for Pro and $0.02 for Fast. These figures do not appear in Cohere's changelog.

Returns only the key parts to cut costs

The new "Context Focus" feature returns the parts of a document most relevant to the question, with character positions and scores. Passing only those parts to generative AI reduces the text it has to read and keeps costs down. It can also be used for highlighting that shows which part of a search result matched. It is turned off by default.

The scores indicating relevance were adjusted so they use the same scale even across different questions and fields. A single threshold score can be used to decide which documents to pass to generative AI. Existing users can switch just by changing the model they specify, without re-ingesting their data. It is available through Cohere's API as well as on the clouds of Microsoft and Amazon.