Cohere’s Embed 5 and Rerank 5 are now one OpenRouter key away

Cohere has made its Embed 5 and Rerank 5 models available on OpenRouter, according to an official changelog entry dated November 4. Developers can now add four Cohere retrieval models to search and RAG pipelines using their existing OpenRouter API key.

Four model IDs, two tiers each

Cohere lists four models on the platform. Embed 5 Pro (cohere/embed-5-pro) is aimed at the best retrieval quality, mainly for offline indexing. Embed 5 Fast (cohere/embed-5-fast) targets low latency and high throughput, such as interactive search, agent loops and heavy query traffic.

On the ranking side, Rerank 5 Pro (cohere/rerank-5-pro) aims for the best ordering quality, especially for difficult queries that demand reasoning and for lengthy or specialist documents. Rerank 5 Fast (cohere/rerank-5-fast) is the quicker, cheaper option for teams where response time and spend matter most. Cohere says it beats Rerank 4 Pro.

Embed 5 accepts text, images and mixed inputs such as PDF pages. Cohere also says the Pro and Fast embedders share an embedding space, so a corpus can be indexed with Pro while Fast handles the query side. That split matters in practice: the expensive pass runs once offline, and the cheap, quick model handles every live query.

Pricing is not in the changelog. Cohere points readers to each model’s page on OpenRouter. The company also links a guide on using its models through OpenRouter, which covers a setup checklist and the differences from the Cohere API, and OpenRouter’s cookbook for a retrieval pipeline with embeddings and reranking.

What the models already offered elsewhere

The OpenRouter listing does not repeat the specifications, but Cohere’s own announcements do. Its Embed 5 post describes a 128k-token context, support for more than 100 languages, and Matryoshka embeddings at sizes from 256 up to 2048 dimensions, with float, int8 and binary output types. Cohere says the model shows “major gains over Embed 4” on visually rich documents, financial filings, parsed PDFs, code and multilingual retrieval, though that post gives no benchmark numbers.

Rerank 5, covered in our earlier article, offers a 32k-token context window, more than 100 languages and an optional Context Focus mode that returns the relevant spans inside each document. Until now both families were available through Cohere’s own API, Microsoft Foundry and Amazon SageMaker, with dedicated deployment via Model Vault.

A routing layer that shortens the path to Cohere

The angle here is competition for developer attention. Embeddings and reranking have mostly been bought one vendor at a time, each with its own account, billing and key. OpenRouter is best known for routing chat models through one endpoint, and adding Cohere’s retrieval models lets a team test them next to other providers without new procurement.

This fits Cohere’s wider push. In October the company was reported to be raising a round of CAD 7 billion at a valuation of $35.7 billion, with the Canadian government as an investor, as we reported. Distribution channels such as OpenRouter, Foundry and SageMaker help a company known for enterprise deployments reach developers who never talk to a sales team.

Rivals are also pressing on retrieval. Nvidia’s Nemotron 3.5 release put out five open retrieval models, including an 8B embedder, which can be run on a team’s own hardware. Cohere’s models are hosted services, so the trade is convenience against control.

For existing Rerank users, the other detail that matters is billing. Third-party price guides list Rerank 4 per search, and one tracker says Rerank 5 moved to per-token pricing. Neither point is confirmed in Cohere’s changelog, so check the OpenRouter pages before comparing costs.

What the changelog leaves unanswered

The entry is short and gives no benchmark figures, prices, rate limits or regional details. The claim that Rerank 5 Fast beats Rerank 4 Pro is Cohere’s own and has not been independently verified in the material we reviewed. The guide Cohere links describes differences from the Cohere API, so features available on Cohere’s endpoint, such as Context Focus or specific embedding sizes, may not map one to one on OpenRouter. Teams should confirm that against the setup checklist before migrating production traffic.