# Cohere Embed 5 and Rerank 5 on OpenRouter: Pricing and API Details

Cohere's retrieval models now fit into an existing OpenRouter account, but a shared API does not remove the decisions inside a search pipeline.

By Paras Madan | 2026-11-05


Cohere announced on November 4 that Embed 5 and Rerank 5 are available through OpenRouter. Developers can use the models with their existing OpenRouter key and credits, bringing document embeddings, reranking and answer generation into one billing setup. The models themselves are not new releases.

OpenRouter's listings date their availability to November 3. The [Cohere announcement](https://cohere.com/blog/embed-rerank-5-openrouter) explains how to connect them into a retrieval-augmented generation pipeline, where an application searches its own documents before asking a language model to answer.

## Embed 5 and Rerank 5 pricing on OpenRouter

The listed prices below are in US dollars per million input tokens. Embedding image input has a separate rate.

| Model | Model identifier | Text input price |
| --- | --- | --- |
| [Embed 5 Pro](https://openrouter.ai/cohere/embed-5-pro) | cohere/embed-5-pro | $0.12 |
| [Embed 5 Fast](https://openrouter.ai/cohere/embed-5-fast) | cohere/embed-5-fast | $0.08 |
| [Rerank 5 Pro](https://openrouter.ai/cohere/rerank-5-pro) | cohere/rerank-5-pro | $0.05 |
| [Rerank 5 Fast](https://openrouter.ai/cohere/rerank-5-fast) | cohere/rerank-5-fast | $0.02 |

Both Embed 5 variants list image input at $0.40 per million image tokens. These are model charges, not the total cost of running document search. Vector storage, retrieval infrastructure and the model generating the final answer remain separate costs.

All four listings currently say the models are unavailable through OpenRouter's EU in-region routing. A team with an EU processing requirement should resolve that constraint before moving a workload.

## What changes in the retrieval workflow

Embedding turns documents and questions into vectors that can be compared for semantic similarity. Reranking then scores a smaller set of retrieved documents against the question. The answer model receives the selected evidence.

This integration reduces account and billing work for a team already using OpenRouter. It does not create the document index or decide which evidence belongs in a prompt. Those choices still sit in the application.

Embed 5 supports text, images and mixed text-image input, with a 128,000-token input limit. Its output dimensions range from 256 to 2,048. Pro and Fast share an embedding space, so documents indexed with Pro can be queried with Fast when the dimensions match.

That gives teams a useful split to test: spend more on indexing documents, then use the cheaper model for repeated user queries. It is not a general rule that embeddings from different model families can be mixed.

## The limits to check before switching

Rerank 5 accepts text and returns documents ranked by relevance. Its query-and-document window is about 32K tokens, displayed as 33K on OpenRouter. Adding it also adds another request before answer generation, which matters for interactive search.

A better ranking can reduce the amount of irrelevant text sent to the answer model. But filtering too aggressively can remove an exception or qualification needed for a correct response. Compare the final answers, not just the ranking scores.

For an existing search product, the useful experiment is small: keep the questions and candidate documents fixed, compare Pro and Fast, then measure answer quality, latency and total cost. A single API makes that experiment easier to operate. The result still depends on whether the right evidence reaches the reader.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn launches and complex product updates into clear stories that developers and founders want to read and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).