Breaking01 / AI
Cohere Embed 5 and Rerank 5 join OpenRouter on existing keys
One API key can now run a full retrieval pipeline, but the announcement gives no per-token prices.

Cohere announced on 5 Nov 2026 that OpenRouter users can call Embed 5 and Rerank 5 with their existing API keys. They use the same credits and need no separate setup.
Call structure. Indexing for RAG (retrieval-augmented generation, a method that finds relevant documents and uses them as grounding before answering) runs once per document collection, according to Cohere's announcement. Each question then takes one call each to Embed 5 Fast, Rerank 5 and a chat model.
Cohere suggested 1,024-dimension int8 vectors (int8 stores each value as an 8-bit integer) as the best point of quality against cost. OpenRouter's embeddings API returns floating-point values, and int8 and binary vectors are available through Cohere's own API. Rerank 5 reads up to 32k tokens at a time, counting the query and documents together.
Scoring. Cohere said a Rerank 5 score carries the same meaning across queries and domains. In the company's example on fictional documents, items with embedding scores of 0.337 and 0.316 fell below a 0.3 cutoff after Rerank 5 (company-published figures).
Cohere recommended indexing with Embed 5 Pro and querying with Embed 5 Fast. The two models share an embedding space, so they can be mixed as long as both sides use the same dimensions. In tests across 40 datasets, that pairing kept 98.4% of the retrieval quality of an all-Pro setup (company-published figure).
Fast on both sides kept 96.6% in the same tests (company-published figure). With any other model family, Cohere wrote, both sides must use the same model.
Cohere did not disclose per-token OpenRouter prices for the two models in the announcement.