A search assistant may retrieve the right document and still pass far more text to its language model than the answer requires. Cohere’s Rerank 5, released on 21 October, adds a way to select relevant passages inside those documents. Its Context Focus option returns scored character spans alongside document-level rankings, letting an application choose the evidence it sends onwards.[1][2]
The release has two tiers: Pro prioritises ranking quality, while Fast targets higher-volume, lower-latency workloads. Both fit after an initial keyword, vector or hybrid search. Developers can retain their existing collection and retrieval system because reranking reads the candidate text at query time.[2]
The agent test measured the complete retrieval stack
Cohere evaluated a search agent across 17 question-answering datasets. Starting with BM25 keyword search, accuracy was 46%. Adding Rerank 5 Pro raised it to 49%. With Context Focus set to a 0.4 threshold, the full stack reached 58% accuracy and used 35% fewer total tokens per task than BM25 alone. These are supplier-reported results for its specified agent and evaluation setup.[1]
The total-token measure includes input, output and reasoning across the run, including cached tokens. That makes it broader than the length of a single retrieved passage. The accuracy gain from 46% to 58% is 12 percentage points; the token saving uses the original keyword-only setup as its baseline.[1]
Where each stage contributes
| Stage | Job | Relevant measurement |
|---|---|---|
| Initial retrieval | Find candidate documents | Recall, latency and search cost |
| Reranking | Order candidates for the query | Ranking relevance |
| Context Focus | Select relevant spans inside documents | Context retained and tokens removed |
| Answer generation | Use the selected evidence | Answer accuracy, citations and total cost |
A faster tier and a straightforward upgrade path
In separate retrieval tests, Cohere says Pro ranked first on 39 of 46 composite datasets and in the top two on 43. It reports 82.2 on BRIGHT, 12.4 points above Rerank 4 Pro, using Cohere’s RCP-nDCG@10 evaluation method. These measure document ranking, whereas the agent experiment measures answers produced downstream.[1]
Fast processed about 187 documents of roughly 1,000 tokens each per second on a single NVIDIA A100 in Cohere’s load test, 2.9 times Pro’s throughput. Launch API prices were US$0.05 per million tokens for Pro and US$0.02 for Fast. Cohere lists API, Model Vault, Microsoft Foundry, Amazon SageMaker and private deployment options.[1]
An upgrade changes the model parameter to rerank-v5.0-pro or rerank-v5.0-fast. Context Focus is optional and off by default; when enabled, it returns offsets into the original document so the application can preserve the connection between a selected passage and its source.[2]
The design gives developers a useful control over how much evidence reaches an answer. The practical balance is to remove irrelevant material while retaining the sentence that supplies an exception, condition or necessary explanation. Passage selection makes that balance explicit rather than leaving it to the size of a document chunk.
Sources & further reading
- Introducing Rerank 5: precise, efficient enterprise document retrieval Cohere · 21 Oct 2026 · Accessed 25 Oct 2026
- Announcing Cohere’s Rerank 5 Models Cohere Developer Docs · 21 Oct 2026 · Accessed 25 Oct 2026