A company may process an archive once and search it millions of times. Using the same embedding model for both jobs ties an occasional ingestion expense to a recurring query expense. Cohere’s latest release allows those decisions to come apart without rebuilding the archive’s vector index each time the query model changes.
Two models announced on September 30
Cohere released Embed 5 on September 30, 2026, in Pro and Fast tiers. Both convert text, images or combined text-and-image inputs into vectors used to retrieve relevant material. They support more than 100 languages and a context window of 128,000 tokens.
Pro targets retrieval quality; Fast targets lower cost and latency. Listed text-embedding prices are $0.12 per million tokens for Pro and $0.08 for Fast. Image pricing is separate. The models are available through the Cohere API and other deployment channels, with support for private serving.
The shared coordinate system is the release’s hinge
Embedding models usually need compatible document and query representations. If their coordinates do not line up, a query vector cannot reliably identify the closest relevant documents. A smaller query model is therefore useful only if it still speaks the index’s numerical language.
Pro and Fast share that language. Cohere recommends indexing with Pro and querying with Fast for many customers. The expensive document representation stays in place while a cheaper model handles repeated live requests. Both sides must use the same output dimension; Cohere says compatibility also holds with supported dimensional truncation and int8 quantization.
Cohere tested the corpus/query combinations across 40 development datasets. With Pro documents and Pro queries normalized to 100, Fast queries against Pro documents scored 98.4. Pro queries against Fast documents scored 97.3. These are normalized average retrieval scores, not percentages of user questions answered correctly.
That detail puts a price on compatibility. Sharing an embedding space avoids a mandatory rebuild, but it does not make every pairing identical. The gap needs to be measured on the customer’s corpus, particularly for rare terminology and documents whose layout carries meaning.
Some headline scores measure reranking
The release reports results across financial, multilingual and multimodal document collections. A methodological footnote matters: results using Cohere’s RCP-nDCG@10 metric compare models reordering a fixed candidate set. Those results measure reranking quality, rather than the complete first-stage search problem.
A model can be good at sorting supplied candidates while still missing useful material during an initial search. Cohere says it evaluates first-stage retrieval separately with nDCG and recall. Readers comparing the release’s charts need to keep the stage and metric attached to each number.
The distinction becomes practical in an enterprise pipeline. A better vector representation cannot recover a relevant page that ingestion discarded, or compensate for a candidate-generation step that never surfaces it. End-to-end answer quality needs its own evaluation.
The index can dominate the economics
Model calls are only one expense. Embed 5 supports shorter vectors and lower-precision outputs, letting operators exchange some representational capacity for storage and search efficiency. Cohere recommends 1,024-dimensional int8 vectors for many deployments and describes binary vectors as an option for a compact first pass before reranking.
Raw vector size is not the same as the complete database bill. Metadata, indexing structures, replicas and query infrastructure also consume resources. An evaluation should hold those operational choices visible while comparing retrieval quality, query latency and the cost of refreshing a changing archive.
For a large, slowly changing corpus with heavy query traffic, the shared-space design offers a distinct operating choice: spend more when documents arrive and less each time someone asks about them. For a rapidly changing collection, ingestion speed may matter more. Embed 5 gives those workloads separate controls, making the balance between reading and asking something an operator can measure rather than inherit.
References
Questions, news tips or corrections? Contact [email protected].
