Cohere released Embed 5 on September 30. The model IDs are embed-v5.0-pro and embed-v5.0-fast.
The release notes say the two models share an embedding space. A corpus indexed with one model can be queried with the other.
So the model that embeds documents and the model that embeds queries can be chosen separately. Cohere recommends Pro for indexing and Fast for queries.
Vector compatibility set by output dimension and input type
The launch announcement attaches a condition: document vectors and query vectors must use the same output dimension.
Cohere says compatibility also holds with Matryoshka truncation, which shortens vectors, and int8 quantization, which stores values at lower precision. That describes compatibility of vector representations, not identical search results.
The embeddings documentation says to set input_type="search_document" for text being searched over and input_type="search_query" for search queries. Vector length is chosen with the output_dimension parameter.
If these settings are not recorded when comparing models, there is no way to tell whether a difference came from the model or the settings.
Caveats: Embed 4 indexes and self-reported figures
The shared-space compatibility statement covers the two Embed 5 models. It is not evidence that an earlier Embed 4 index can be reused as is.
Cohere says it tested cross-model pairings on 40 development datasets, with average losses of 1.6% for Fast queries and 2.7% for Pro queries.
These are Cohere’s own figures. Compatibility alone is not enough to prove retrieval quality on a specific corpus.
A query-side comparison with fixed conditions
The next step is a controlled query-side comparison. Fix the Pro-indexed documents, output dimension, vector format, search settings and representative queries, then compare Fast and Pro on retrieval relevance and end-to-end latency.
Include difficult queries in the languages and domains that matter to the application.
If Fast meets the requirements, the documented shared space allows switching only the query path without regenerating document vectors. This is an evaluation plan, not a test performed for this article.
A shared-space guarantee saves an index rebuild, but retrieval quality on your own queries is a value that still has to be measured.