Cohere releases Embed 5 Pro and Fast with a shared embedding space

Cohere shipped Embed 5, a multimodal embedding family whose Pro and Fast variants share one vector space, so you can index with one and query with the other.

Cohere has released Embed 5, an embedding family in two variants: embed-v5.0-pro for quality-critical offline indexing and embed-v5.0-fast for low-latency queries and agent loops. For RAG builders, the useful detail is that Pro and Fast share an embedding space. A corpus indexed with one can be queried with the other, and Cohere recommends indexing with Pro and querying with Fast.

Per Cohere's changelog, both variants embed text, images and mixed inputs such as PDF pages as a single vector. They support 100+ languages and a 128k-token context. Matryoshka outputs come in 256 to 2048 dimensions as float, int8 or binary. Embed 5 is available now through the Embed API, Microsoft Foundry, Amazon SageMaker and, for single-tenant deployment, Model Vault.

AlphaSignal reports list prices of $0.12 per million tokens for Pro and $0.08 for Fast. Those figures aren't confirmed on Cohere's own pages. It also relays Cohere's figures that cross-model pairings scored 1.6% and 2.7% below same-model baselines. Per the same write-up, Cohere measured Fast at 2.4x Pro's document throughput, but no method is stated.

The benchmark leads come from Cohere's own evaluation: HuggingNews repeats the claim that Pro tops Gemini Embedding 2, Voyage 4 Large and Jina Embeddings v5 on ViDoRe V3. No independent evaluation has surfaced yet.

Several questions remain open. It's unclear how much quality you lose with int8 or binary vectors at low dimensions. It's also unclear whether Embed 4 vectors are compatible, so plan for a possible re-index. The shared-space idea isn't unique to Cohere: a recent comparison notes Voyage's voyage-4 family also produces interchangeable embeddings. Test on a labelled sample of your own corpus before switching.

Sources