Cohere Rerank 5 ships in two tiers, Pro and Fast, and adds an option called Context Focus that also returns scored passages from inside each document, alongside the usual document-level score. Cohere says both models can be used today through its own API and the Model Vault hosting service, on Microsoft Foundry and Amazon SageMaker, and to all users of its North AI platform.
API pricing is $0.05 per million tokens for Pro and $0.02 for Fast. Teams that want dedicated capacity can rent Model Vault instances instead, and Cohere prices a medium instance of either Rerank 5 model at $4 an hour, or $2,500 a month.
With Context Focus on, a single call hands back a score for each document plus scored snippets from inside it, and the request controls how strict the cutoff is and how many snippets come back. Cohere ran it across 30 question-answering datasets, with Qwen3.5-9B writing the answers and Qwen3.5-35B grading them. Dropping half of the retrieved text left 96 percent of the answer quality intact, and dropping three-quarters left 87 percent. In a search-agent test, keyword search followed by the Pro model with Context Focus at a 0.4 threshold used 35 percent fewer tokens than keyword search alone. The share of correct answers rose from 46 to 58 percent.
Cohere says reranker scores until now were mostly good for sorting, so teams had to tune a cutoff for each model and tune it again after every upgrade. Rerank 5 returns a relevance score meant to read as a probability: 0.7 means a 70 percent chance the document is useful. Cohere checked that against blind human graders and NIST's TREC Deep Learning assessors, reporting an expected calibration error of about 0.06 on both. Rerank 4 Pro scored 0.384 and 0.309 on the same two sets.
The even-odds line sits at 0.50 for Pro and 0.45 for Fast. Cohere says one threshold can then replace a fixed number of top results: a developer passes along every document above it, and an agent can rephrase its query or say it does not know when no document clears the bar.
Cohere's ranking claims rest on RCP-nDCG@10, a variant of the standard metric that it developed itself. On BRIGHT, a reasoning-heavy retrieval benchmark, Pro scores 82.2. That is 12.4 points above Rerank 4 Pro and 5.8 above Voyage AI Rerank 3. Cohere says Pro ranked first on 39 of the 46 composite datasets it evaluated.
Fast beats Rerank 4 Pro on all four headline benchmarks. On a single NVIDIA A100 it reranks about 187 documents of 1,000 tokens each per second, which Cohere puts at 2.9 times Pro's throughput.
The reranker sits on top of existing keyword, vector or hybrid search without re-indexing. The Rerank documentation lists the model IDs as rerank-v5.0-pro and rerank-v5.0-fast, and installs inside a customer's own VPC or on-premises go through Cohere's sales team.
