Cohere released Embed 5 on September 30, 2026: a family of two embedding models, Embed 5 Pro and Embed 5 Fast, that share one vector space. Pro costs $0.12 and Fast $0.08 per million text tokens on Cohere's API, both accept text and images with a 128K-token context, and a corpus indexed with one can be queried with the other. The models are commercial: Cohere's announcement offers them through its API, cloud marketplaces and private deployments, and does not announce a public weights release.
Embedding models turn documents and queries into vectors for search and retrieval-augmented generation (RAG). Cohere's announcement positions Embed 5 for complex enterprise content: financial filings, parsed PDFs, visually rich documents where meaning sits in tables, charts and layout, code, and more than 100 languages.
The Cohere changelog entry for September 30, 2026 names the two model IDs and their intended roles:
embed-v5.0-pro for the highest retrieval quality, especially offline indexing.embed-v5.0-fast for low latency and high throughput, such as interactive search, agent loops and high-volume queries.Because Pro and Fast produce compatible vectors, Cohere recommends indexing with Pro and querying with Fast. That splits the cost and latency profile of a retrieval system in a useful way: the expensive, quality-sensitive step (embedding the corpus) happens once and offline, while the frequent, latency-sensitive step (embedding each user query) uses the cheaper, faster model. It also means a team can change its query-time model without re-indexing, which is normally the most expensive part of switching embedders.
| Embed 5 Pro | Embed 5 Fast | |
|---|---|---|
| Text price (per 1M tokens) | $0.12 | $0.08 |
| Image price (per 1M tokens) | $0.40 | $0.40 |
| Context length | 128K tokens | 128K tokens |
| Inputs | Text, images, fused text and image | Text, images, fused text and image |
| Output dimensions | 2048, 1536, 1024, 768, 512 or 256 | Same |
| Embedding formats | float, int8, binary | Same |
Source: Cohere announcement, September 30, 2026, checked October 10, 2026.
Matryoshka representations and int8 or binary output let teams trade some retrieval quality for much smaller vector indexes, which matters at enterprise scale. Cohere says Embed 5 is generally available on the Cohere API and Model Vault, Microsoft Foundry and Amazon SageMaker. Its pricing page also lists dedicated Model Vault instances for both models, starting at $3.00 per hour or $2,000 per month for a small instance.
Cohere lists self-hosting as supported for both tiers and says quantized weights lower serving requirements for private deployments. That is private deployment under a commercial agreement, through Cohere's on-premises and isolated-VPC offerings, not an announced open-weight release. For teams that need an embedding model they can download and run under a standard open-source license, EmbeddingGemma 2, released by Google under Apache 2.0 on October 6, 2026, is the comparable release in this period.
Cohere says Embed 5 Pro achieved the highest average score of any model it tested, with particular strength on financial documents, parsed PDFs and visually rich documents. On the ViDoRe V3 document-retrieval benchmark, it reports 85.8 for Pro and 84.5 for Fast, against 83.7 for Voyage 4 Large, 83.2 for Gemini Embedding 2 and 75.5 for OpenAI's text-embedding-3-large. It also reports first place for Pro on three financial benchmarks (FinanceBench, FinQA and the ViDoRe V3 finance subset), with Fast second on each.
Two cautions apply. First, these are Cohere's evaluations, and this article does not reproduce them. Second, Cohere scores Embed 5 with a new methodology of its own, RCP-nDCG@10, which judges retrieved documents against query-specific relevance criteria rather than only fixed labels. That may well be a better measure, but it means the numbers are not directly comparable with leaderboard results computed the traditional way.
Enterprise RAG teams with document-heavy corpora, particularly financial filings, scanned or parsed PDFs, and mixed text and image pages.
High-traffic search systems that can index with Pro and query with Fast to cut query latency and cost.
Organizations that need private deployment but are comfortable with a commercial license, through Model Vault, Foundry, SageMaker or Cohere's on-premises offerings.
When was Cohere Embed 5 released?
September 30, 2026.
How much does Embed 5 cost?
$0.12 per million text tokens for Pro and $0.08 for Fast; images cost $0.40 per million tokens on both, per Cohere's announcement.
Can I query an Embed 5 Pro index with Embed 5 Fast?
Yes. Cohere says the two models share an embedding space and recommends indexing with Pro and querying with Fast.
Is Embed 5 open source?
Not as announced. Cohere offers it through its API, cloud marketplaces and private deployments under commercial terms, and its announcement does not describe a public weights release.
Discover more content: