Cohere launches Rerank 5 with passage-level Context Focus

Cohere launched Rerank 5 on Oct. 21, 2026, in Pro and Fast tiers. The optional Context Focus feature returns scored passages from retrieved documents in the same rerank call. Cohere said both models are generally available, giving enterprise search and retrieval-augmented generation systems a way to send selected passages instead of entire documents to a downstream language model.

Reranking is the second stage of retrieval. After keyword, vector or hybrid search produces candidate documents, another model orders them by relevance. Context Focus narrows the output further. When an application enables the relevant_spans field, it can set a score threshold and a maximum number of passages. Qualifying results include a score and character start and end offsets. The feature is off by default, according to Cohere’s release notes.

The model identifiers are rerank-v5.0-pro and rerank-v5.0-fast. Cohere positions Pro for maximum ranking quality on complex or domain-specific queries, while Fast targets lower-latency, higher-throughput workloads. The company says existing retrieval indexes do not need to be rebuilt because its reranker reads raw text at query time. Users can upgrade by changing the model parameter.

The benchmark results have not been independently reproduced in the opened sources. Using its RCP-nDCG@10 methodology, Cohere said Pro ranked first on 39 of 46 composite datasets and at least second on 43 across ViDoRe V3, nanoMTEB, BRIGHT and BEIR. The company reported a BRIGHT score of 82.2, 5.8 points above Voyage AI Rerank 3 and 12.4 points above Rerank 4 Pro, plus smaller leads over the strongest external competitor it tested on ViDoRe V3, BEIR and nanoMTEB.

For Fast, Cohere reported throughput of about 187 documents of roughly 1,000 tokens each per second on one NVIDIA A100, or 2.9 times Pro’s throughput, with nearly three-times-faster response times. The company said the test used 96 documents per request, concurrency of eight and more than 1,000 requests for each document length.

Cohere also said Context Focus reduced the text passed downstream while retaining most answer quality in its tests. Across 30 question-answering datasets, removing half of the retrieved context lowered answer quality by two percentage points, to 96% of the baseline. Removing three-quarters lowered it by seven points, to 87%. In a separate 17-dataset search-agent evaluation, a 0.4 threshold cut token use by 34% compared with Rerank 5 Pro without Context Focus and reached 58% accuracy. Cohere said the full system used 35% fewer tokens than BM25 alone while scoring 58% versus 46%. Across six deep-search benchmarks, token use fell 41% while accuracy rose from 39% to 52%.

Cohere’s models documentation identifies Rerank 5 Pro and Fast deployments on Amazon SageMaker and Microsoft Foundry. Cohere says the models are also available through its API and Model Vault, and to all North users. An AWS Marketplace listing separately shows Rerank 5 Fast as a Cohere-sold SageMaker model and documents the optional Context Focus request field.