Cohere Launches Rerank 5 Pro and Fast with Context Focus for Agentic RAG

Cohere releases Rerank 5 Pro and Fast, featuring Context Focus to identify relevant passages within documents and reduce context size by up to 10x.
In agentic deep-search benchmarks (6 datasets), the full stack (BM25 + Rerank 5 Pro + Context Focus) reduces token usage by 41% and increases accuracy from 39% to 52% compared to BM25 alone. Across the broader set of 17 datasets, token usage fell by 35%.
Rerank 5 Pro outperforms competitors on 39 of 46 complex datasets, scoring 82.2 on BRIGHT and leading in finance, medical, and scientific domains.
The models exhibit significantly better calibration, with an Expected Calibration Error (ECE) approximately 4x lower than Voyage AI Rerank 3 and Rerank 4 Pro.

Performance and Efficiency

Rerank 5 Pro scores 82.2 on BRIGHT, exceeding Voyage AI Rerank 3 by 5.8 points and Rerank 4 Pro by 12.4 points.
Rerank 5 Fast processes approximately 187 documents per second on a single NVIDIA A100 GPU, offering 2.9x higher throughput and 3x faster response times than Pro.
Pricing is set at $0.05 per 1M tokens for Pro and $0.02 per 1M tokens for Fast.

Context Focus allows agents to remove irrelevant text spans; at a threshold of 0.4, it reduces tokens by 34% while raising accuracy to 58% in QA tasks.
Calibrated relevance scores enable consistent filtering across domains, with Pro's default threshold for "likely useful" documents at 0.50.
Rerank 5 models achieve an ECE of 0.057-0.058, compared to 0.226-0.384 for competing models, indicating more accurate probability estimates for document utility.

Rerank 5 Pro and Fast are generally available via Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker, and North.
The models integrate with existing search stacks (keyword, vector, hybrid) without requiring re-indexing and support LangChain, Haystack, Elasticsearch, and Weaviate.