Cohere Brings Embed 5 and Rerank 5 to OpenRouter’s Unified API

Cohere announced on November 4, 2026 that its Embed 5 and Rerank 5 retrieval models are available on OpenRouter. Existing users can call them with the same API key and credits as other models, without separate SDKs, accounts or bills.

The models cover the retrieval stages of a retrieval-augmented generation, or RAG, pipeline: Embed 5 converts documents and queries into vectors, Rerank 5 rescores candidates against the query, and any chat model on OpenRouter can generate the final answer. Cohere also published a Python reference workflow spanning indexing, retrieval, relevance filtering and cited generation.

OpenRouter’s Embed 5 Pro listing prices the model at $0.12 per million text tokens and $0.40 per million image tokens, with Embed 5 Fast listed at $0.08 per million tokens. Its Rerank 5 Pro listing prices that tier at $0.05 per million input tokens and Rerank 5 Fast at $0.02 per million tokens.

The OpenRouter rollout expands distribution rather than introducing another model generation. Cohere introduced Embed 5 on September 30, 2026, and Rerank 5 on October 21, 2026; both families offer Pro and Fast tiers.

Shared Embeddings Split Indexing and Query Costs

The central architectural detail is that Embed 5 Pro and Fast share an embedding space. A corpus can therefore be indexed with Pro and queried with Fast without rebuilding the index, provided both sides use the same output dimensions. That separates indexing work, which runs once per corpus, from the latency- and cost-sensitive embedding call made for every question.

In Cohere’s tests across 40 development datasets, normalized mean retrieval quality was 98.4 for a Pro-built index queried with Fast and 96.6 with Fast on both sides, against 100 for Pro on both sides. The OpenRouter workflow adopts that Pro-index/Fast-query pairing, reserving Pro for the offline indexing stage.

Embed 5 offers six vector sizes from 256 to 2,048 dimensions, with 2,048 as the default. The OpenRouter example uses 1,024-dimensional floating-point vectors because its embeddings API returns floats; Cohere separately recommends int8 vectors at that dimension as its preferred quality-to-cost point. Batching chunks reduces network round trips during indexing but does not reduce token-based embedding charges. The online path uses one Embed 5 Fast call, one Rerank 5 call and, when context survives the filter, one chat call per question.

Relevance Scores Control Generation Context

The reference pipeline retrieves 20 chunks by cosine similarity, sends them with the question to Rerank 5, keeps as many as five results and discards scores below a demonstration threshold of 0.3. Reranking reads the query and each candidate together rather than relying only on vector proximity, and can operate after semantic, keyword or hybrid search without Embed 5.

Cohere says Rerank 5’s calibrated scores remain consistent across queries and domains, allowing one threshold to decide which context reaches a generator; its launch documentation describes a score of 0.7 as a 70% probability that a document is useful. Pro is intended for reasoning-heavy queries and long or domain-specific documents, while Fast targets high-volume search and agents that search repeatedly. The models support a 32K-token context and can process JSON records or tables formatted as YAML strings.

At generation, surviving chunks become numbered sources, with instructions to answer only from them and cite them inline. If none clears the threshold, the workflow reports that it could not find the answer and skips the chat-model call. Because Rerank reads raw text at query time, switching reranker versions requires no re-indexing or re-embedding.

Demonstration Results and Production Constraints

Cohere’s demonstration uses fictional API documentation and asks how to rotate a key without downtime. Embedding-only retrieval ranked the correct Rotating API keys document first at 0.522, followed by Webhook signing secrets at 0.337 and Creating API keys at 0.316. After reranking, the correct document scored 0.808 and both near-matches fell below the cutoff.

The sample requires Python 3.10 or newer, an OpenRouter API key and Python SDK, and NumPy for its in-memory vector store. Production implementations are directed to replace that store with a vector database. OpenRouter’s RAG guidance presents reranking as a precision-versus-latency choice: it can improve results when a broad candidate set must be narrowed, but adds an API call and is less suitable when minimum latency is the priority.

Multimodal retrieval needs a workaround. Embed 5 accepts images and mixed text-and-image inputs such as PDF pages, and OpenRouter’s embeddings API accepts image URLs, but Rerank 5 is text-only. The proposed options are to rerank parsed page text or captions, or omit reranking for visual results.

OpenRouter lists Cohere as the sole provider for both Pro models and forwards requests directly to it. Both listings state that EU in-region routing is unavailable.

The OpenRouter availability consolidates the retrieval stack’s integration and billing, while its production economics remain tied to the architecture: one-time Pro indexing, recurring Fast query, rerank and chat calls, and the relevance threshold that controls generation context. The single API does not erase the added latency, text-only reranking or regional-routing constraints documented for that stack.