Cohere has launched Rerank 5, a new suite consisting of two models: Pro, designed for reasoning-heavy tasks and long documents, and Fast, optimized for high throughput and low latency. Both models support over 100 languages and a 32K context window, allowing for the processing of semi-structured data like JSON, tables, and YAML without requiring re-indexing.
The release introduces Context Focus, an optional feature that identifies and returns specific relevant spans within documents. According to Cohere's testing on 30 QA datasets, this feature can reduce context usage by up to 10 times while maintaining significant answer quality. In search-agent tests, the use of Context Focus at a 0.4 threshold reduced token consumption by 34% while increasing accuracy to 58%.
Rerank 5 provides calibrated probability scores, achieving an expected calibration error (ECE) of approximately 0.06. This performance surpasses previous versions and competitor models like Voyage AI Rerank 3. The Pro model leads across major benchmarks, including ViDoRe V3, BEIR, BRIGHT, and nanoMTEB, while the Fast model outperforms several competitors in speed-focused metrics.
The models are available via the Cohere API, Model Vault, Microsoft Foundry, and Amazon SageMaker. API pricing is set at $0.05 per 1 million tokens for the Pro version and $0.02 per 1 million tokens for the Fast version. Model Vault instances are priced at $4.00 per hour. Integrations are supported for platforms including LangChain, Haystack, Elasticsearch, and Weaviate.
