Cohere Releases Rerank 5 With Context Focus and Calibrated Scores

The model ships as rerank-v5.0-pro, aimed at ranking quality on reasoning-heavy queries and long, domain-specific documents, and rerank-v5.0-fast, aimed at low latency and high throughput for agents running many searches. Both support a 32K token context window and more than 100 languages, and both can rerank semi-structured data such as JSON, tables and YAML records. Cohere said upgrading requires no re-indexing; developers change the model parameter and keep existing keyword, vector or hybrid search. Cohere did not disclose parameter counts and offers no open weights or quantized builds.

Context Focus is optional and off by default. A single Rerank call returns document scores plus scored character spans, each with start and end offsets into the original text; developers set it through the relevant_spans field with a threshold and max_spans value. Cohere tested it on 30 QA datasets and said removing half the retrieved text kept 96% of answer quality, while removing three quarters kept 87%. Cohere said teams can push to 10 times less context at some quality cost.

In a search-agent test across 17 datasets, BM25 alone scored 46% accuracy, Cohere said. Adding Rerank 5 Pro raised it to 49%; Context Focus at the default 0.1 threshold cut tokens per task by 19% and reached 57%; at 0.4, tokens fell 34% and accuracy hit 58%. Against BM25 alone, the full stack used 35% fewer tokens. On six deep-search benchmarks, tokens fell 41% and accuracy rose from 39% to 52%.

Rerank 5's calibrated scores read as probabilities: a score of 0.7 means a 70% chance the document is useful. Cohere measured expected calibration error against blind human graders and NIST TREC-DL assessors and reported about 0.06 on both. Voyage AI Rerank 3 scored 0.234 and 0.307, while Rerank 4 Pro scored 0.384 and 0.309, according to the report. Cohere said developer teams can filter by threshold instead of a fixed top-k, and an agent can detect when nothing relevant came back.

All benchmark scores are self-reported by Cohere using RCP-nDCG@10, a Cohere-developed variant of nDCG, and Cohere ran the competitor models itself; the results have not been independently verified. Pro leads Cohere's four main benchmarks with 92.2 on ViDoRe V3, 90.4 on BEIR, 82.2 on BRIGHT and 87.3 on nanoMTEB, and ranks first on 39 of 46 composite datasets in Cohere's own evaluation. BRIGHT shows the largest jump over Rerank 4 Pro, at 12.4 points. Fast beats Voyage AI Rerank 3 Lite, Qwen3-Reranker-8B and Rerank 4 Pro on all four, ties Voyage AI Rerank 3 on ViDoRe V3 at 90.6 and trails it on BEIR, 88.1 versus 88.6. On parsed PDFs averaged over 47 datasets, Pro scored 89.9, Fast 88.5 and Voyage AI Rerank 3 88.4. Pro also ranks first on 14 of 19 non-English languages in Cohere's internal suite.

Rerank 5 is generally available on the Cohere API and Model Vault, and runs on Microsoft Foundry and Amazon SageMaker for both tiers, as well as private VPC and on-premises deployments. API pricing is $0.05 per 1 million tokens for Pro and $0.02 for Fast; Model Vault lists $4.00 per instance-hour for both tiers on its Medium tier. Integrations include LangChain, Haystack, Elasticsearch and Weaviate, according to Cohere's integration docs. Cohere benchmarks speed on a single NVIDIA A100 and reported median latency of 11.7 seconds per 96-document request for Pro, versus 4.1 seconds for Fast.

Editor's Summary

Cohere released Rerank 5 in Pro and Fast tiers, adding passage-level Context Focus and calibrated scores that are intended to reduce tokens sent to LLMs. Cohere's self-reported benchmarks show Pro ahead on its main datasets, and API pricing is $0.05 per 1M tokens for Pro and $0.02 for Fast. The results have not been independently verified.