A new provider route can simplify access. It does not settle which evidence your final answer needs.
Cohere announced Embed 5 and Rerank 5 on OpenRouter on November 4. Developers can access the models with an OpenRouter API key. This adds a distribution route to the Rerank 5 release we covered in October.
OpenRouter adds Pro and Fast variants
Cohere lists cohere/embed-5-pro, cohere/embed-5-fast, cohere/rerank-5-pro, and cohere/rerank-5-fast. It positions Pro toward quality-sensitive workloads and Fast toward lower latency and higher throughput. Those product descriptions still need evaluation against the intended corpus.
Embed 5 Pro and Fast share an embedding space. Cohere says a team can index documents with Pro and embed queries with Fast. That compatibility belongs to the embedding models; it should not be confused with a claim that every provider route has identical pricing or API behavior.
The release note directs readers to the individual OpenRouter model pages for prices and to its setup guide for differences from the Cohere API. Record the route and billed model in a comparison rather than assuming that changing a provider is only a credential change.
Context Focus changes what the answer model sees
Cohere’s October 21 Rerank 5 release introduced rerank-v5.0-pro and rerank-v5.0-fast. Context Focus can return relevant passages as character ranges within supplied documents. It is disabled unless requested.
Reranking orders documents at query time, so upgrading the reranker does not itself require rebuilding embeddings. Selecting only passages makes another change: the downstream model receives less context. A highly relevant paragraph can still omit the exception in the next section.
Evaluate the ranking and answer separately
Begin with fixed queries, versioned documents, and known source locations. Compare ranking while retaining full document context. Then enable passage selection and examine whether the final answer still includes required exceptions, table footnotes, or conflicting evidence.
Save the returned order and character ranges alongside the final answer. Count unsupported answers and human corrections, as well as latency and total retrieval-plus-generation cost. A relevance score alone cannot prove the answer is correct.
If passage selection drops a necessary qualifier, test a wider passage or retain the full document for that question type. Our FinRank report explains a related problem: an answer can look right while citing the wrong evidence. These are proposed checks; MustHave has not benchmarked the new OpenRouter route.



