Cohere releases Rerank 5 to select the passages AI systems read

Cohere released Rerank 5 Pro and Fast on October 21, 2026, giving developers a new way to control what evidence reaches an AI system. Both models can identify relevant passages inside documents, not just rank whole documents. The optional feature targets the amount of text an AI must process before it answers.

A reranker sits after an initial search and reorders its results by relevance. Rerank 5 works with existing keyword, vector or combined search systems without requiring documents to be indexed again. Its new Context Focus feature adds a finer choice: which parts of those results are worth sending onward.

Selecting evidence within a document

Context Focus returns document relevance scores and scored text spans in one API call. Developers can adjust a threshold to retain more evidence or remove more text. The application can then pass selected passages to the language model, rather than treating every retrieved document as equally necessary reading.

Cohere tested that tradeoff across 30 question-answering datasets. Each system selected five documents from 50 candidates, then removed text at different thresholds. One language model answered from the remaining material, and another graded the answers. Cohere reports that removing half the context reduced answer quality by two percentage points; removing three-quarters reduced it by seven.

A separate company test put retrieval inside a search agent. Across 17 datasets, adding Pro to keyword search raised accuracy from 46% to 49%. With Context Focus set to 0.4, accuracy reached 58%, while total tokens per task fell 35% against keyword search alone. That count included input, output, reasoning and cached tokens across the agent’s calls—not just the passages it retrieved.

Scores that can also filter

Rerank 5 also introduces calibrated relevance scores: numbers intended to reflect the probability that a document is useful, rather than merely its position in a ranking. Cohere says developers can use those scores to filter results by relevance instead of always sending a fixed number of documents to the answering model.

That creates another option when search comes back empty of useful evidence. If nothing clears the threshold, an agent can rewrite its query or say it does not know, rather than answer from whichever result ranked first. Cohere checked calibration against a blind human-grading panel and NIST search assessors, reporting an average gap of about 0.06 between scores and observed usefulness.

Choosing quality, speed and deployment

Cohere reports that Pro finished first on 39 of 46 evaluated datasets and at least second on 43. Those results use RCP-nDCG@10, the company’s evaluation method designed to address incomplete relevance labels in search benchmarks. They are Cohere’s measurements using its own methodology, rather than an independent assessment of the release.

Fast emphasizes throughput. In Cohere’s tests on a single NVIDIA A100 GPU, it processed about 187 documents of roughly 1,000 tokens per second—2.9 times Pro’s throughput. The company also reports nearly three times faster response times. The test used 96 documents per request and eight concurrent requests, making the workload part of the comparison.

Access routes

• Both models are generally available through the Cohere API and Model Vault.
• Microsoft Foundry and Amazon SageMaker also offer the models, and access extends to all North users.
• For deployment in a customer’s own virtual private cloud or on-premises, Cohere directs teams to contact the company.