Cohere announced on November 4, 2026, that Embed 5 and Rerank 5 were available through OpenRouter. Existing OpenRouter users could access them with their current API key and credits, without a separate setup, according to Cohere’s announcement. The change brings Cohere’s retrieval models to another API platform.
In a retrieval pipeline, embeddings find candidate material; reranking sorts those candidates before an answer is generated.
How Embed 5 and Rerank 5 fit into a retrieval workflow
Retrieval-augmented generation, or RAG, fetches relevant source material before an AI model writes an answer. Embed 5 turns content into vectors, numerical representations that let a system find material related to a query. Rerank 5 scores candidate documents against that query and orders them by relevance. A chat model can then use the strongest matches to generate a response.
| Component | Role in the workflow | What happens next |
| Embed 5 Pro | Creates vectors for indexed content; its OpenRouter listing describes text, image, and mixed text-image inputs. | Embed 5 Fast can encode a search query in the same embedding space when the vector dimensions match. |
| Embed 5 Fast | Creates a vector for the incoming query. | The system retrieves candidate documents for reranking. |
| Rerank 5 | Scores and reorders text candidates against the query. | The application can pass selected material to a chat model. |
| Chat model | Generates an answer using the selected material. | The response is grounded in retrieved sources. |
Cohere’s example retrieves 20 candidates, reranks them, keeps up to five, and uses a demonstration score threshold of 0.3 to decide what reaches the generation step. If no candidate clears that threshold, the example skips the chat-model call. Cohere also says Rerank 5 can follow an existing semantic, keyword, or hybrid search system; Embed 5 is not a prerequisite for reranking. The workflow and example are described in Cohere’s OpenRouter post.
For image-based retrieval, Cohere describes Embed 5 Pro as accepting image and mixed text-image input, while Rerank 5 is text-only. Cohere’s suggested options are to rerank parsed text or captions, or to leave image results out of the reranking step.
What Cohere reports about Context Focus
In a post published October 21, 2026, Cohere described Context Focus as an optional Rerank 5 feature that selects relevant text spans and returns them with document-level relevance scores. In an evaluation across 30 question-answering datasets, Cohere reported that answer quality remained at 96% after half the retrieved text was removed, and at 87% after three quarters was removed. Those figures describe Cohere’s reported evaluation; they are not a guarantee for a particular application.
Cohere’s October post describes Context Focus as part of Rerank 5, while its November 4 announcement covers the models’ availability through OpenRouter. The two announcements address different details of the retrieval stack. Cohere’s Rerank 5 post describes the feature and evaluation.
Routing for the listed Pro models
OpenRouter’s listings state that EU in-region routing is unavailable for Embed 5 Pro and Rerank 5 Pro. The restriction applies to those listed Pro models; it does not establish the routing status of the Fast tiers. See the Embed 5 Pro listing and the Rerank 5 Pro listing.
