The short answer
Cohere is a Canadian AI company, founded in Toronto in 2019, that builds models and software for businesses rather than consumers. Its flagship model is Command A+, released on 20 May 2026 with open weights under the Apache 2.0 licence, and its enterprise platform is North 2, announced on 5 October 2026, which gives staff AI agents, search and automations over company data, deployed in the company's own cloud, on premises or fully air-gapped.
| Made by | Cohere, Canada (founded in Toronto in 2019) |
|---|---|
| Free plan | A free trial API key for testing only; no consumer app, and trial keys cannot be used commercially |
| Models | Command A+ (flagship), Command A family, North Mini Code, North Small Translate, Embed 5, Rerank 5, Cohere Transcribe, Tiny Aya |
| Best at | Private AI for regulated companies, search over company documents, multilingual work |
| Shows sources | North grounds answers in company knowledge; how citations appear depends on the setup |
| Team plans | North and Enterprise, with custom pricing through sales |
| Your data | Paying API customers can switch off training in Data Controls; private deployments send Cohere no prompts or outputs |
What Cohere is in late 2026
Cohere does not make a chat app you download on a Sunday afternoon. It makes the engines and the workplace platform that companies run for their own staff. If your employer tells you to use an internal AI assistant that answers questions from company files, it may be built on Cohere, and knowing what sits underneath helps you use it well.
There are two halves to the business. The first is the models, available through the Cohere API and as open weights. The flagship is Command A+, released on 20 May 2026. It is a sparse mixture of experts model with 218 billion total parameters, 25 billion of them active at a time, a 128K token input context, up to 64K tokens of output, image input and support for 48 languages. Cohere says it runs on as few as two H100 graphics cards when quantised, and it describes Command A+ as the last model in the Command A family. Around it sit specialist models: Embed 5 and Rerank 5 for search, Cohere Parse for turning documents into structured text, Cohere Transcribe for speech, North Mini Code for agentic coding and North Small Translate for translation across more than 50 languages.
The second half is North, the platform staff actually see. North 2, announced on 5 October 2026, adds a redesigned agent system plus Skills, Libraries, Memory, Applications (which turn a chat into decks, dashboards, documents or small apps), Automations and Connectors. It works with Cohere's models or a company's own choice of model, and it can be deployed self-hosted, in a private cloud, hybrid, on premises or completely air-gapped.
What Cohere is best at, with real examples
Answering questions from your company's own documents
Search is where Cohere has always been strongest. Embed 5 and Rerank 5 are built to find the right passage in a large pile of contracts, policies or support tickets, and North uses that to ground answers in company knowledge. A practical example: a compliance analyst asks North which supplier contracts mention a data transfer clause, and gets a list with the passages, rather than searching SharePoint folder by folder.
Keeping data inside the building
Banks, hospitals, public bodies and manufacturers often cannot send documents to a public AI service. Cohere's answer is to run where the data already lives. North 2 connects to Slack, SharePoint, OneDrive, Microsoft Outlook, Microsoft Exchange, Jira, Linear, Notion and GitHub today, with PitchBook, Crunchbase, S&P Global and FactSet among the connectors Cohere lists as planned. Cohere cites SOC 2 Type 2, ISO 27001 and ISO 42001 certifications.
Work in many languages
Command A+ handles 48 languages and comes with a tokeniser that Cohere says cuts token counts by 20% for Arabic, 16% for Korean and 18% for Japanese. That matters for teams that write or search in more than one language every day.
Running models yourself
Because Command A+, North Mini Code and Cohere Transcribe are released under Apache 2.0, a technical team can download them from Hugging Face and run them on its own hardware. Note the exception: North Small Translate's open weights are licensed for non-commercial use only, under CC BY-NC 4.0.
Where Cohere falls short
- It is not for individuals. There is no consumer plan. If you want an assistant for your own work, ChatGPT, Claude or Gemini are the practical choices, and our comparison tool helps you pick.
- You cannot just sign up for North. North is sold through sales with custom pricing. You can request a demo or try an instant demo, but a full rollout needs IT involvement.
- The free key is for testing. Every account gets a rate-limited trial API key, and Cohere's pricing page says trial keys may not be used for production or commercial purposes.
- Answers are only as good as the connected data. If the right SharePoint site is not connected, North cannot find the document, and it may not tell you what it cannot see unless you ask.
How to get better results from Cohere North
- Name the source. Say which library, folder or system to search, such as "search the HR policies library only". Narrow searches return more precise passages.
- Ask for the passages, not just the answer. Request the exact quote and the document name for every claim. This makes checking quick and catches answers built on the wrong file.
- State the output shape. North can build tables, charts, documents and decks. Say "a table with supplier, clause and renewal date" instead of "tell me about our suppliers".
- Turn repeat questions into Skills or Automations. If you ask the same weekly question, ask your North admin to save it as a reusable Skill or a scheduled Automation, so everyone gets the same answer format.
- Ask what it could not check. End important requests with "list anything you could not find or could not access". Gaps in access are the most common reason for a confident but incomplete answer.
Privacy settings and where to find them
For paying customers on Cohere's hosted platform, you can opt out of having prompts and generations used for training at any time. The setting is in the Cohere Platform under Settings, then Data Controls in the left-hand menu: set the toggle to Off. The change applies only to data created after you switch it. Logged prompts and generations on the hosted platform are deleted automatically after 30 days, with exceptions for legal requirements, contracts or suspected misuse, and approved customers can get zero data retention.
The bigger privacy point is deployment. In private deployments and on third-party cloud platforms, Cohere says it does not receive any prompts or generations at all. Enterprise use is governed by the customer's agreement and data processing addendum, so your company's contract, not a toggle, is what really decides where your data goes.
Which Cohere option suits whom
Pick North if you are an IT or data leader in a regulated company that needs AI agents over internal knowledge without sending documents to a public service. It is not for a ten-person agency that just wants better emails.
Pick the API and open models if you are a developer building search, retrieval or multilingual features, especially if you want to self-host Command A+. Start with the free trial key for experiments and apply for a production key through the Billing and Usage page before anything goes live.
Skip Cohere if you are an individual professional looking for a personal assistant. You will get more from a general assistant, and our Cohere lesson and AI stacks show where Cohere fits in a company setup. Check Cohere's pricing page for current terms, since enterprise pricing is set case by case.
Prompts to copy
Find every contract with a specific clause
Search only the [library or folder name] library. List every contract that contains a clause about [topic, e.g. data transfer outside the EU]. For each one give the document name, the exact clause text, the counterparty and the renewal date, in a table. At the end, list any documents you could not open.
You get: A checkable table of clauses with exact quotes and a list of gaps
Answer a policy question with the source
Using our HR policies only, answer this question from an employee: [question]. Quote the exact policy passage you rely on, name the document and its last updated date if shown, and say clearly if the policies do not cover the question.
You get: A policy answer you can trust because the passage is quoted
Prepare a customer briefing
I meet [customer name] on [date]. Search our Outlook email, Jira tickets and Slack messages about them from the last 90 days. Write a one-page briefing: open issues, recent decisions, people involved and three questions I should ask. Link every point to its source.
You get: A sourced one-page briefing before a customer meeting
Compare two versions of a document
Compare [document A] with [document B]. List every change in obligations, dates, amounts and responsibilities in a table with columns for section, old wording, new wording and why it matters. Ignore formatting changes.
You get: The meaningful changes between two versions, without the noise
Draft a weekly automation
Every Monday at 9:00, search [team channel or folder] for decisions, risks and deadlines mentioned in the last seven days. Produce a short summary with three sections (Decisions, Risks, Deadlines next 14 days), each point linked to its source, and send it to [team].
You get: A repeatable weekly summary you can ask your admin to save as an Automation