Cohere Rerank 5 highlights the passages that matter and returns calibrated scores

What's it about?

Cohere has released Rerank 5, a new family of reranking models. A reranker re-sorts search results before they go to a large language model (LLM). That matters for RAG (retrieval-augmented generation: a model gets matching documents and answers based on them) and for AI agents that run many searches. Cohere's pitch: send fewer, but better, tokens to the LLM.

Rerank 5 comes in two tiers: rerank-v5.0-pro and rerank-v5.0-fast. Cohere doesn't disclose the parameter count. Both have a 32K context window, support more than 100 languages, and can also rank semi-structured data such as JSON, tables and YAML. Switching doesn't require reindexing: you change the model parameter and keep your existing keyword, vector or hybrid search.

Context Focus: relevant passages instead of whole documents

Context Focus is optional and off by default. A single rerank call returns document scores plus scored character spans with start and end positions in the original text. You control it through the relevant_spans field, with a threshold and max_spans.

Cohere tested this on 30 QA datasets: when half of the retrieved text was removed, 96% of answer quality was retained; with three quarters removed, 87%. According to Cohere, up to 10x less context is possible with some loss in quality.

In a test with search agents across 17 datasets, BM25 (classic keyword search) alone reached 46% accuracy. With Rerank 5 Pro, it was 49%. With Context Focus at the default threshold of 0.1, tokens per task dropped 19% at 57% accuracy; at 0.4, they dropped 34% at 58%. Compared with BM25 alone, the full system used 35% fewer tokens. On 6 deep-search benchmarks, tokens fell 41% while accuracy rose from 39% to 52%.

Calibrated scores

Rerank 5's scores are meant to be readable as probabilities: 0.7 means a 70% chance that the document is useful. Cohere measured the expected calibration error (ECE, lower is better) against blind human raters and NIST TREC-DL assessors. Rerank 5 sits at around 0.06 on both, Voyage AI Rerank 3 at 0.234 and 0.307, and Rerank 4 Pro at 0.384 and 0.309. In practice, this lets you filter by threshold instead of a fixed top-k, and an agent can tell when nothing relevant was found.

Benchmarks (vendor claims)

All figures use RCP-nDCG@10, a variant of nDCG developed by Cohere. Cohere ran the competing models itself; nothing has been independently verified.

On BRIGHT, Pro is 12.4 points ahead of Rerank 4 Pro. Fast beats Voyage Rerank 3 Lite, Qwen3-Reranker-8B and Rerank 4 Pro on all four benchmarks, ties Voyage Rerank 3 on ViDoRe V3 (90.6) and trails it slightly on BEIR (88.1 vs. 88.6). On parsed PDFs (47 datasets), Pro scores 89.9, Fast 88.5 and Voyage Rerank 3 88.4. According to Cohere, Pro leads in 14 of 19 non-English languages and ranks first on 39 of 46 datasets in its own evaluation.

Availability and pricing

Rerank 5 is generally available through the Cohere API and Model Vault, as well as on Microsoft Foundry and Amazon SageMaker, and for all North users. Private VPC and on-prem deployment are also offered. The API costs $0.05 (Pro) and $0.02 (Fast) per 1M tokens; Model Vault lists $4.00 per instance hour in the Medium tier for both. Integrations include LangChain, Haystack, Elasticsearch and Weaviate, among others. There are no open weights or quantized versions. Cohere measures speed on a single NVIDIA A100: median latency is 11.7 s per request with 96 documents for Pro and 4.1 s for Fast.

Our take

For teams with RAG or agent pipelines, Context Focus is the most interesting part, because it can cut token costs without requiring you to build a second extraction stage. Calibrated scores make it easier to set fixed thresholds across different domains. The caveats: all figures are self-reported, use an in-house metric, and Cohere tested the competition itself. Pro's latency of almost 12 seconds per request is high for interactive applications. If you need open weights, you'll have to stick with alternatives like Qwen3-Reranker-8B (Apache 2.0), though it scores lower in Cohere's tests.