What Is Cohere? Enterprise LLM Provider

While OpenAI courts consumers with ChatGPT and Anthropic positions Claude as the thoughtful generalist, Cohere has built a business doing something narrower and, for its customers, more useful: selling large language models exclusively to enterprises. Founded in 2019 and headquartered in Toronto, Cohere counts Oracle, Salesforce, Notion, RBC, and LG among its customers, was valued at $6.8 billion in an August 2025 funding round, and is led by a co-author of the paper that invented the transformer architecture in the first place.

This article explains what Cohere actually builds, why its enterprise-only strategy exists, how its models compare to the household names, and when a technical buyer should pick it over an OpenAI or Anthropic deployment.

Founders and origin story

Cohere was founded in June 2019 by three researchers who had recently worked at or alongside Google Brain, the machine-learning research group inside Google that helped popularize the modern deep-learning era.

• Aidan Gomez (CEO) — One of the eight co-authors of "Attention Is All You Need," the 2017 Google Brain paper that introduced the transformer architecture that now powers essentially every modern large language model. Gomez was a 20-year-old undergraduate intern at Google Brain when he co-wrote it, the youngest author on the paper. He later completed a PhD at the University of Oxford as a Clarendon Scholar, pausing partway through to launch Cohere and finishing the degree in 2024.
• Nick Frosst — Trained under Geoffrey Hinton, one of the "godfathers of deep learning," working on capsule networks (a neural network design intended to better preserve spatial relationships between features than standard convolutional networks). Frosst also fronts a band called Good Kid.
• Ivan Zhang — Met Gomez through For.ai, an open machine-learning research collective, before the three formalized Cohere as a company.

From day one, the founders bet that enterprises would want large language models without the consumer baggage: no chatbot brand to defend to the public, no viral hallucination screenshots to apologize for, and strong data-residency and compliance guarantees baked in from the start. Think of it as the difference between a car manufacturer that sells directly to drivers and a company that supplies engines to a dozen other car brands — Cohere chose to be the engine supplier, not the dealership.

That bet has aged well financially. Cohere's valuation moved from roughly $2.1 billion in 2022 to $5.5 billion by the third quarter of 2024, and then to $6.8 billion in an August 2025 round led by Radical Ventures and Inovia Capital, with Nvidia, AMD Ventures, PSP Investments, and Salesforce Ventures also participating, according to CNBC's reporting on the round. A follow-on extension in September 2025 pushed the figure to roughly $7 billion, per BetaKit. Alongside that round, Cohere hired Joelle Pineau, Meta's former VP of AI Research, as Chief AI Officer, and Francois Chadwick, formerly CFO at Uber and Shield AI, as CFO — a sign the company was staffing up for a more aggressive enterprise push rather than a consumer pivot.

Aidan Gomez was the youngest of the eight authors on "Attention Is All You Need," and he co-wrote it as a Google Brain intern before he had even finished his undergraduate degree. That paper has since been cited well over 100,000 times and underlies GPT, Claude, Gemini, Llama, and Cohere's own models alike.

The model lineup: four families, four jobs

Cohere does not chase a single do-everything flagship the way OpenAI or Anthropic do. It ships four distinct model families, each aimed at a specific enterprise workload, and expects customers to combine them rather than pick just one.

Command: text generation and agentic tool use

Command is the chat and instruction-following family, and the one most directly comparable to GPT or Claude.

• Command R (August 2024) — A roughly 32–35B-parameter model tuned for retrieval-augmented generation (RAG, a technique that lets a model look up real documents before answering — more on that below) and multi-step tool use, released with open weights under a non-commercial research license.
• Command R+ — A larger sibling built for the same RAG- and tool-heavy workloads at higher quality.
• Command A (March 2025) — Cohere's current flagship, a 111-billion-parameter model with a 256,000-token context window that, according to VentureBeat's coverage of the launch, runs on just two GPUs and delivers roughly 150% higher throughput than Command R+. It is tuned specifically for enterprise-relevant tasks: SQL generation, code translation, tool use, and multilingual agent workflows.
• Command A Vision (July 2025) — Added multimodal (image-plus-text) understanding to the Command A line, per Cohere's own changelog.

Embed: turning text into searchable vectors

An embedding model converts a piece of text into a list of numbers (a "vector") that captures its meaning, so that a computer can measure how similar two pieces of text are just by comparing those number lists. Cohere's Embed family has historically ranked among the strongest options on the Massive Text Embedding Benchmark (MTEB), the standard public leaderboard for this kind of model. Embed v4, released in late 2025, produces 1,024-dimensional embeddings, supports variable-length outputs through a technique called Matryoshka Representation Learning (which lets a single model produce shorter, cheaper vectors when full precision isn't needed), and is priced at roughly $0.01 per million tokens, among the more competitive rates in the category. Multilingual versions cover more than 100 languages. Embeddings are the backbone of any RAG pipeline, since a system has to find the right documents before an LLM can read them.

Rerank: fixing bad search results before the model sees them

Rerank is a narrower, cheaper model that takes a first-pass list of search results (say, the top 100 candidate passages an embedding search returned) and reorders them by true relevance to the question asked. Rerank 3.5 is Cohere's current version. In practice, this is often the single cheapest fix available for a struggling retrieval system — swapping in a good reranker can meaningfully improve answer quality without touching the generation model at all.

Aya: multilingual and open-weight research models

Aya Expanse (8B and 32B parameter versions) covers 23 languages, including Arabic, Hindi, Vietnamese, and Indonesian, and is released under a permissive license that allows commercial fine-tuning. It comes out of Cohere Labs (formerly Cohere For AI), the company's nonprofit research arm, and represents Cohere's most open contribution to the field — a deliberate choice to build goodwill and research credibility in languages the larger labs have historically underserved.

What Cohere is good at

1. Multilingual performance. Cohere invested in non-English performance earlier and more deliberately than most competitors. Aya Expanse and the multilingual Embed and Command models consistently hold up better on lower-resource languages than similarly sized English-first alternatives, because Cohere Labs built dedicated multilingual training pipelines rather than treating other languages as an afterthought.

2. Retrieval-augmented generation. Command R, R+, and A were designed around RAG from the start: they cite sources cleanly, follow grounded-answering instructions, and stay closer to the provided context instead of inventing details — the core failure mode known as hallucination. If your product is "answer questions about our internal documents accurately," Cohere's models are tuned for exactly that job in a way general-purpose chat models aren't by default.

3. On-premises and sovereign deployment. Cohere will deploy inside a customer's virtual private cloud on AWS, Azure, Oracle Cloud Infrastructure, or on bare-metal Nvidia hardware. It is one of the few frontier-adjacent providers offering data residency in Canada, the EU, and the Middle East out of the box, which matters enormously for banks, government agencies, and healthcare systems that are legally barred from sending data across borders.

4. Agentic deployment through North. In August 2025, Cohere took its North platform (first piloted with select clients in January 2025) to general availability. North is a low-code workspace where enterprises build custom AI agents — automated assistants that can search, summarize, and generate reports across connected tools like Google Drive, Slack, and Salesforce — and run them entirely inside their own infrastructure, including air-gapped environments with no internet connection at all, according to TechCrunch's report on the launch. North directly competes with Microsoft 365 Copilot and ChatGPT Enterprise, but leads on the same pitch Cohere has made since 2019: keep sensitive data inside the customer's own walls.

5. Actual enterprise procurement. Real contracts, real SLAs (service-level agreements), and the SOC 2, ISO 27001, and HIPAA paperwork that procurement and compliance teams require before they'll sign off on a vendor. That sounds unglamorous, but for a regulated bank or hospital system, it's often the deciding factor over which model scores two points higher on a benchmark.

What Cohere is not good at

• Brand recognition. Most consumers have never heard of Cohere, and that's by design — but it means a product depending on end-user trust in a familiar AI brand should look elsewhere.
• Frontier reasoning benchmarks. Cohere does not chase the top spot on the hardest math and coding leaderboards the way OpenAI's o-series, Anthropic's Claude Opus line, and Google's Gemini do. Command A is competent and fast, not state of the art on raw reasoning.
• Community size. The Hugging Face, Reddit, and YouTube ecosystem around Cohere is smaller than around OpenAI or open-weight Llama models. Stack Overflow answers and third-party tutorials skew heavily toward OpenAI's API conventions.
• No consumer product. There is no Cohere chatbot app to point a curious executive at, which occasionally makes internal buy-in harder even when the technical case is strong.

Comparison: Cohere vs OpenAI vs Anthropic vs Google for enterprise

Table: How the four major enterprise LLM providers stack up

Dimension	Cohere	OpenAI	Anthropic	Google (Vertex AI)
Frontier reasoning	Good	Best-in-class	Best-in-class	Best-in-class
RAG out of the box	Excellent	Good	Good	Good
Multilingual coverage	Excellent	Good	Good	Excellent
First-party embeddings	Top-tier (Embed v4)	Top-tier	None	Top-tier
First-party reranking	Yes (Rerank 3.5)	No	No	No
On-prem / air-gapped	Yes, multi-cloud + bare metal	Limited (Azure)	Limited (AWS Bedrock)	GCP only
Data residency options	Canada, EU, Middle East, US, APAC	US, EU	US, EU	Global
Agentic platform	North (GA, Aug 2025)	ChatGPT Enterprise / Operator	Claude Code / Agent SDK	Agentspace
Open weights	Some (Command R, Aya)	None	None	Gemma (smaller models)
Consumer brand	Weak	Strongest	Strong	Strong
Latest valuation	~$7B (Sep 2025)	~$500B+ (2025 rounds)	~$183B (2025)	N/A (Alphabet subsidiary)

Valuations and product availability change quickly in this market; figures reflect the most recent publicly reported rounds as of this writing.

When to pick Cohere

Pick Cohere if you check two or more of these boxes:

• You are a regulated enterprise, such as a bank, insurer, healthcare provider, or government agency.
• You need data residency outside the United States, particularly Canada, the EU, or the Middle East.
• Your product is heavily retrieval-based: internal search, customer support over your own documentation, or document question-answering.
• You operate across multiple languages, especially languages outside the major European and East Asian set.
• You want one vendor covering generation, embeddings, and reranking together, rather than stitching together three different APIs.
• You need agents that run entirely inside your own infrastructure, including air-gapped environments.

Pick OpenAI, Anthropic, or Google instead if you need the strongest possible reasoning on hard technical tasks, you're building a consumer-facing product where brand trust matters, or you want the largest surrounding ecosystem of tutorials, integrations, and third-party tools.

The partnership strategy: be the engine, not the brand

Cohere has largely skipped the "build a flashy consumer app" stage and instead embedded itself directly into other companies' infrastructure and cloud platforms.

• Oracle invested in Cohere and bundled its models directly into Oracle Cloud Infrastructure's Generative AI service, per Oracle's own product documentation.
• Azure, AWS, and Google Cloud all offer Cohere models through their respective model marketplaces, meaning a customer already committed to one of the big three clouds can add Cohere without a separate vendor relationship.
• Salesforce, an investor in Cohere's 2025 round, has integrated Cohere technology into its Einstein AI features.
• Notion uses Cohere for select AI features inside its product.

This "be the engine, not the brand" approach is close to the opposite of OpenAI's early consumer-first playbook, and for the specific segment Cohere targets — large, risk-averse organizations that already have vendor relationships with Oracle, Salesforce, or a major cloud provider — it has proven to be a workable way to get the models embedded without ever having to win a public brand war.

Common confusions about Cohere

"Cohere makes a chatbot like ChatGPT." It doesn't, at least not for the public. Cohere's models are accessed through APIs, cloud marketplaces, and the North agent platform by business customers, not through a free consumer app people download on their phones.

"Cohere is a small player that doesn't matter." Its valuation trajectory says otherwise: from roughly $2.1 billion in 2022 to $5.5 billion in 2024 to roughly $7 billion by late 2025, with Nvidia, AMD, Salesforce, and PSP Investments among its backers. It is smaller than OpenAI or Anthropic in headline valuation, but it occupies a defensible niche those two companies aren't optimized to serve.

"Open-weight models mean Cohere is an open-source company." Only partly true. Command R and the Aya family are released with open weights, but the flagship Command A remains proprietary and is accessed through Cohere's API or private cloud deployments, not downloaded and run for free by anyone.

"Embeddings and reranking are the same thing." They solve adjacent but different problems. An embedding model turns text into a vector so a system can search across millions of documents quickly. A reranker takes the shortlist that search already narrowed down and re-scores it for true relevance before the best few passages ever reach the language model's context window. Cohere is one of the few vendors that sells both as first-party products designed to work together.

The takeaway

Cohere's core insight was that the biggest AI market wasn't consumers chatting with a bot, it was enterprises that needed a model they could deploy inside their own security perimeter, in their own language, with contractual guarantees a consumer product will never offer. By building Command, Embed, Rerank, and Aya as a coordinated toolkit rather than one flagship chatbot, and by embedding itself inside Oracle, Salesforce, and the major clouds rather than fighting for consumer mindshare, Cohere built a $7 billion business without most people ever hearing its name. For a technical buyer weighing providers, the decision isn't really "which model is smartest" — it's whether your workload looks like open-ended reasoning (favor OpenAI, Anthropic, or Google) or like grounded, multilingual, compliance-heavy retrieval at enterprise scale (favor Cohere).

FAQ

Q: Is Cohere open source? A: Partly. Command R and the Aya Expanse models are released under open licenses for research use, and Aya's license permits commercial fine-tuning. The flagship Command A model remains proprietary and is only available through Cohere's API or private deployments.

Q: Can I use Cohere for free? A: Cohere offers a free trial tier on its API for prototyping and evaluation, and the open-weight Command R and Aya models can be downloaded and run without a Cohere subscription. Production use at scale requires a paid plan.

Q: Is Cohere better than OpenAI for RAG? A: For grounded question-answering over your own documents with clean source citations, Command R+ and Command A are tuned specifically for that job and often outperform general-purpose chat models out of the box. For open-ended assistants or creative tasks, GPT or Claude models are generally stronger.

Q: What is the difference between Embed and Rerank? A: Embed converts text into numerical vectors so a system can search semantically across millions of documents. Rerank takes the shortlist of top candidates that search already returned and reorders them by true relevance before the best few reach the language model.

Q: Who is Aidan Gomez? A: Cohere's CEO and co-founder. He co-authored "Attention Is All You Need," the 2017 Google Brain paper that introduced the transformer architecture underlying essentially every modern large language model, at age 20 as an undergraduate intern.

Q: Is Cohere safe for healthcare data? A: Cohere offers HIPAA-compliant deployment options on AWS, Azure, and on-premises infrastructure. As with any vendor, compliance responsibility is shared — a customer still needs a signed Business Associate Agreement and its own proper data-handling practices.

Q: What is Cohere North? A: North is Cohere's low-code agentic AI platform, generally available since August 2025, that lets enterprises build custom AI agents for tasks like search, summarization, and report generation. It can run entirely inside a customer's own infrastructure, including air-gapped environments with no external network access.

Q: How much does Cohere cost compared to OpenAI? A: Cohere does not publish flat consumer pricing the way OpenAI does; enterprise deployments are typically negotiated as custom contracts based on volume, deployment model (cloud API versus on-prem), and support tier. Embed v4, by contrast, does have public API pricing, at roughly $0.01 per million tokens as of its late-2025 release.

Q: How large is Command A? A: Command A has 111 billion parameters and a 256,000-token context window, but it's engineered to run on just two GPUs while delivering roughly 150% higher throughput than its predecessor, Command R+, according to Cohere's own release materials.

Q: Does Cohere compete directly with OpenAI, or is it a different market? A: Both, depending on the workload. On raw chatbot reasoning and consumer reach, Cohere doesn't compete head-on with OpenAI. On enterprise RAG, multilingual support, data residency, and embedded agent deployments inside regulated industries, Cohere positions itself as a direct and, in some cases, better-suited alternative.

Sources

• Cohere: "Cohere Raises $500M to Accelerate Enterprise AI", August 2025
• CNBC: "AI startup Cohere valued at $6.8 billion in latest fundraise, appoints new executives", August 14, 2025
• BetaKit: "Cohere's valuation hits $7 billion USD following $100-million round extension", September 2025
• VentureBeat: "Cohere targets global enterprises with new highly multilingual Command A model requiring only 2 GPUs", March 2025
• Cohere changelog: "Announcing Cohere's Command A Vision Model", July 31, 2025
• TechCrunch: "Cohere's new AI agent platform North promises to keep enterprise data secure", August 6, 2025
• Oracle Cloud Infrastructure documentation: Cohere Command A model
• MTEB (Massive Text Embedding Benchmark) leaderboard, Hugging Face
• Vaswani et al., "Attention Is All You Need," NeurIPS 2017 (arXiv:1706.03762)
• CNBC: "Cohere CEO and ex-Google researcher Aidan Gomez on how AI makes money", July 6, 2024

See also: What Is RAG? · How transformers pay attention to words · What AI agents can and cannot do · AI memory systems explained