Cohere has refreshed the reranking layer of its enterprise search stack. On October 21, 2026 the company released Rerank 5 in two versions: Pro, tuned for ranking quality, and Fast, aimed at latency- and cost-sensitive workloads. Customers can reach both through Amazon SageMaker, Microsoft Foundry, Cohere's own API and its Model Vault, inside Cohere's North workspace product, and for private deployment in a customer's VPC or on premises. Through the API, Fast costs $0.02 per million tokens and Pro $0.05.
A reranker sits after a first-pass search and reorders candidate documents so that a language model or agent reads the most useful ones first. Cohere's pitch for this generation is about the tokens that reach the model. Its new optional feature, Context Focus, returns not just a relevance score per document but scored text spans within it, so an application can forward only the passages that matter. Developers set a threshold to trade context size against how much evidence is kept, and one API call returns both levels.
Cohere's own tests put numbers on that trade. Across 30 question-answering datasets, the company says removing half of the retrieved text kept 96% of answer quality, and removing three quarters kept 87%. In a search-agent setup built on BM25 keyword search with the glm-5-3-flash model, adding Rerank 5 Pro and then Context Focus at a 0.4 threshold lifted accuracy from 46% to 58% while using 35% fewer tokens than BM25 alone, according to Cohere.
The second change concerns what the scores mean. Rerankers usually emit numbers that are good for sorting but shift from model to model, so teams retune cutoffs at every upgrade. Cohere says Rerank 5 scores are calibrated, so a value of 0.7 should correspond to roughly a 70% chance a document is useful. Checked against a blind human panel and NIST TREC Deep Learning judgments, it reports an expected calibration error near 0.06, against 0.23 to 0.38 for Rerank 4 Pro and Voyage AI's rerankers in its comparison.
On ranking quality, Cohere reports that Rerank 5 Pro leads ViDoRe V3, BEIR, BRIGHT and nanoMTEB, ranking first on 39 of 46 composite datasets. Its largest stated gain is on the reasoning-heavy BRIGHT benchmark, where Pro scores 82.2, 5.8 points above Voyage AI Rerank 3 and 12.4 above Rerank 4 Pro. Fast, Cohere says, beats Voyage AI Rerank 3 Lite and both Qwen3-Reranker-0.6B and the larger 8B model, and handles about 187 documents of 1,000 tokens per second on one NVIDIA A100, nearly three times Pro's throughput.
Alongside Embed 5 and Parse 5, Rerank 5 sits in Cohere's current search family and works on top of keyword, vector or hybrid search without re-indexing, with integrations for LangChain, Haystack, Elasticsearch and Weaviate. Cohere also opened a private beta of Compass Cloud, a managed service bundling parsing, embedding, hybrid search and reranking.
The release arrives a day after NVIDIA published open Nemotron 3.5 retrieval models, a sign of how much attention the retrieval layer is getting as agents run more searches per task. All of the figures above are Cohere's. Several results use RCP-nDCG@10, a metric Cohere developed itself, and the multilingual comparison relies on an internal suite, so outside replication will matter before treating the leads as settled.