From the source
Cohere's Embed 5 and Rerank 5 retrieval models are now live on OpenRouter, giving existing OpenRouter users access through a single API with unified billing and no additional SDKs.
- The post walks through building a working RAG pipeline in a few dozen lines of Python, using the OpenRouter Python SDK and NumPy for an in-memory vector store.
- Embed 5 Pro and Fast share one embedding space, so a corpus can be indexed with embed-5-pro for quality and queried with embed-5-fast for latency.
- Rerank 5 scores are calibrated, meaning a score means the same thing across queries and domains, so one relevance threshold can filter weak context before it reaches the LLM.
- Embed 5 returns Matryoshka vectors at six sizes, 2048 by default, and the post suggests 1,024-dim int8 vectors as the best quality-to-cost point.
- Embed 5 accepts up to 128k tokens, but every chunk also passes through Rerank 5, which reads up to 32k tokens of query and chunk combined and splits anything longer into several passes.
