Cohere brings Embed 5 and Rerank 5 to OpenRouter

Cohere announced Embed 5 and Rerank 5 on OpenRouter on November 4, 2026. Developers can use their existing OpenRouter account to add document retrieval and relevance ranking alongside a generation model. Both families have Pro and Fast variants. OpenRouter’s current model listings date their availability to November 3.

The practical change is a new access route for existing models. ByteForward’s coverage of the original Embed 5 launch explains the September release. The OpenRouter integration makes a combined search and answer workflow easier to assemble, but it does not expose every feature available through Cohere directly.

What the four models do

Embedding search represents documents and questions as vectors, then retrieves nearby matches. A reranker examines the question together with each shortlisted document to decide which results deserve priority. A separate language model writes the answer. Keeping those jobs distinct helps identify whether a weak response came from missing evidence, poor ordering or generation.

Embed 5 Pro is the higher-accuracy option for indexing a collection. Embed 5 Fast targets lower-latency queries and larger request volumes. Both accept text, images and mixed inputs, support more than 100 languages and list a 128K-token context window. Six output sizes are supported, from 256 to 2,048 dimensions.

The two Embed variants share a vector space. Teams can index documents with Pro and query that index with Fast, provided both sides use the same output dimension. That separates the recurring query cost from the work of preparing the corpus. A compatible vector space does not remove the need to evaluate retrieval on the documents people actually search.

Rerank 5 Pro is positioned for ranking quality on difficult queries and specialist material. Rerank 5 Fast targets high-volume search and agent workflows. These are text rerankers. Their context limit is 32,768 tokens for each query and document combined, not one shared allowance for an entire candidate list.

Current model pricing

The OpenRouter listings checked on November 5 show US dollar rates per million tokens. Embed 5 Pro costs $0.12 for text input, while Embed 5 Fast costs $0.08. Image tokens cost $0.40 per million for either Embed variant. The Fast discount therefore applies to text, not image inputs.

Rerank 5 Pro lists $0.05 per million input tokens and Rerank 5 Fast lists $0.02. These token-based charges should not be confused with a price per search or per document. Sending a longer candidate set changes the amount of material processed even if the application ultimately displays only a few results.

For scale, 100 million text tokens at the listed embedding rates would cost $12 with Pro or $8 with Fast. Those are calculated model charges, not a complete service budget. Vector storage, reranking, answer generation and any applicable account-level charges remain separate considerations. Use measured token volumes rather than document counts when estimating spend.

Where the OpenRouter API differs

Cohere’s integration documentation says Cohere serves these requests, with OpenRouter forwarding them to the Cohere API. The integration uses OpenRouter request and response formats. One account and billing path can simplify access, but existing Cohere client code still needs the correct OpenRouter model IDs and endpoint calls.

Embedding requests support input, input_type, dimensions and encoding_format. The available output encodings are float and base64. Developers who require int8 or binary embeddings must use Cohere’s Embed API. Multimodal input follows OpenRouter’s format rather than assuming the two services accept identical payloads.

OpenRouter’s rerank interface supports query, documents and top_n. It returns each result’s original index, relevance score and document. Context Focus, including relevant_spans, is not part of this API. Applications that need the model to identify relevant spans inside a document still need the Cohere Rerank API.

A workable retrieval setup

Cohere’s example separates offline indexing from the online question path. It indexes with Pro, embeds questions with Fast, reranks a candidate set and sends the selected evidence to a chat model. That is a useful starting architecture, not a production benchmark or a guarantee of accurate answers.

The OpenRouter embedding SDK reference distinguishes search_document from search_query input types. Make that choice explicit when building the two sides of retrieval. Record the model and dimension settings with the index so a later configuration change does not silently mix incompatible representations.

The rerank SDK reference exposes top_n as the number of results to return. Treat candidate selection and final selection as separate tuning decisions. Evaluate whether the first stage finds the evidence at all before trying to improve the ordering of its results.

Cohere describes Rerank 5’s scores as calibrated across queries. That makes a common relevance cutoff a useful option to test. It should still be checked on answerable questions, ambiguous requests and cases with no supporting document. An empty result needs an explicit application response rather than an invitation for the generator to improvise.

Check regional requirements before migrating

All four current OpenRouter listings say they are not available through EU in-region routing. That restriction matters independently of model quality or price. Teams with a regional processing requirement should confirm an eligible deployment route before sending their document collection.

OpenRouter’s regional routing documentation says its shared-endpoint regional routing fails when no eligible endpoint can serve a request, rather than silently leaving the region. Its regional feature also separates prompt processing from account and operational records. A model appearing in the general catalog is therefore not proof that every residency requirement is covered.

For an existing OpenRouter application, the integration is a practical way to trial a retrieval stack without adding another billing account. The decision should turn on the whole workflow: evidence quality, latency, model charges, regional eligibility and whether compact embeddings or Context Focus are required. The cheapest individual call is only one part of that comparison.

Featured image is an original AI-generated conceptual illustration of document retrieval and selection, not a product interface.