Cohere AI Updates: October 10, 2026

1. Cohere Publishes Sizing Guide for Shared vs. Dedicated Embed and Rerank Inference

Cohere. Cohere published guidance on choosing between consumption-based shared inference and provisioned dedicated capacity for its Embed and Rerank models, arguing that request shape and traffic pattern matter more than requests per minute. The post shows rerank cost growing with candidate count, from about 4,400 tokens per search at 20 candidates to about 22,000 at 100, and gives illustrative break-even points against a single NVIDIA A10 instance, such as roughly 20 requests per minute for catalog embedding and about 29 requests per minute for reranking 50 documents. It recommends that many search applications use dedicated capacity for steady catalog indexing and consumption pricing for bursty query traffic. Source