Embed 5

An embedding model turns a piece of content, a paragraph, an image or a whole page of a PDF, into a list of numbers called a vector, so that items with similar meanings end up close together. Search engines and AI assistants use those vectors to find the passages that answer a question before a language model writes the reply. Embed 5 is Cohere’s new family of them, released on 30 September 2026 in two versions: Pro, at $0.12 per million text tokens, and Fast, at $0.08.

Cohere calls it “a new family of frontier embedding models” in its launch post. Both versions read text, images and pages that mix the two, in more than 100 languages, with a context of 128,000 tokens.

It reads text, images and whole pages

Embed 5 takes text, images and mixed text-and-image inputs up to 128,000 tokens long, in more than 100 languages, and returns a vector in one of six sizes, from 256 to 2,048 numbers. The model IDs are embed-v5.0-pro and embed-v5.0-fast, according to Cohere’s changelog. Each vector can come back as standard floating-point numbers, as 8-bit integers or as single bits, which trades a little accuracy for a lot of storage.

The storage gap is large. A 2,048-number vector of 32-bit floats takes 8 KB, and a 256-number binary vector takes 32 bytes, 256 times less, Cohere says.

Pro costs $0.12 per million tokens and Fast $0.08

On the Cohere API, Embed 5 Pro costs $0.12 per million text tokens and Embed 5 Fast $0.08, on Cohere’s pricing page read on 2 October 2026, and images cost $0.40 per million tokens on both, according to the launch post. Indexing a million pages of about 500 tokens each, 500 million tokens, would cost $60 with Pro and $40 with Fast.

Model Vault is Cohere’s dedicated hosting, where a customer rents a whole instance. On Amazon SageMaker, Embed 5 is sold through AWS Marketplace by the hour of the server that runs it.

Cohere’s tests put Pro first on document search

On Cohere’s own run of ViDoRe V3, a benchmark of searching documents such as reports and slides, Embed 5 Pro scores 85.8 and Fast 84.5, ahead of Voyage 4 Large at 83.7 and Google’s Gemini Embedding 2 at 83.2. Cohere ran it on the documents’ parsed text and scored it with RCP-nDCG@10, a measure it introduced the same day. Cohere’s previous model, Embed 4, scores 77.0 on the same run, and OpenAI’s text-embedding-3-large 75.5.

Cohere’s other results lean towards business documents. On FinanceBench Pro scores 80.1 and Fast 80.0; on FinQA, 90.0 and 88.8. On Cohere’s suite of PDFs parsed into text, Pro averages 84.8 against 83.6 for Voyage 4 Large and 80.8 for Gemini Embedding 2, and on pages that fuse text and images, 82.3 against Gemini’s 61.3. RCP-nDCG@10 is scored by “a calibrated AI judge”, and Cohere calls Embed 5 “the first model family evaluated with RCP-nDCG@10”.

Gemini leads on most Asian languages

In Cohere’s own table of ten languages from Japanese to Thai, Google’s Gemini Embedding 2 beats Embed 5 Pro in nine, and Voyage 4 Large beats it in five, by YFarmX’s count of Cohere’s figures. The languages are Japanese, Chinese, Korean, Arabic, Farsi, Hindi, Bengali, Telugu, Indonesian and Thai. Cohere’s claim of the highest multilingual average rests on five European languages, where Pro averages 77 against 76 for Voyage and 73 for Gemini.

Fast gets through 2.4 times as many documents

Embed 5 Fast processes documents at an average of 2.4 times the throughput of Pro, Cohere says, 377.3 documents a second against 159.7 in its throughput chart. Cohere suggests using both: index the documents once with Pro and run the searches with Fast. In its tests that pairing kept 98.4% of the quality of using Pro for both.

Where to get it

Embed 5 is on the Cohere API and Cohere’s Model Vault, on Microsoft Foundry and on Amazon SageMaker, Cohere’s launch post says. Microsoft announced both models in Foundry on the same day, where its catalogue lists them as a preview. Cohere sells private deployments for customers who want to run the models in their own environment.

Strongest on business documents

Embed 5’s best results in Cohere’s own tests are on finance documents, parsed PDFs and pages that mix text and images. For Cohere’s chat models, which answer from the passages an embedding search finds, see Command A+.