Cowork, Work, Vibe Work: Four ways to delegate work to AI

State of affairs updated on September 20, 2026 (clarification on Dispatch on September 27, 2026), based on the official documentation of Anthropic, OpenAI, Microsoft, and Mistral, reviewed page by page. This article expands upon our comparison of August 27 by adding Vibe Work by Mistral. The first version dated from September 5th; in fifteen days, two of the four editors changed things that we presented as settled.

A year ago, we "chatted" with an AI. Today, the four major platforms offer a mode where you no longer ask a question but describe a result. At ChatGPT, the switch is called "Chat | Work," at the top of the page. At Microsoft, "Chat | Cowork At Mistral, the interface is "Chat | Work | Code," and the developer has already announced that the Chat mode is slated to disappear. At Claude, the switch also existed until September 16th; since then, on the Pro and Max plans, there's only one Claude who decides when a request warrants a task with tools, files, and an action plan. The process remains the same everywhere: the AI ​​plans, opens your tools, generates files, asks for your approval before acting, and returns with a deliverable.

Behind this apparent convergence, the four publishers have made very different choices. Here's what truly distinguishes them, on the points that matter when you want to entrust work to an agent.

What has changed in the last two weeks

• anthropic merged the cat and Cowork September 16th ("Claude Cowork is now just Claude"), launched Claude Docs and Claude Slides in beta, released the Fable 5.1 template on September 1, and integrated a browser into the desktop application panel on August 26.
• OpenAI released GPT-6 Astra on September 3, created $100 "Premium" Business seats on August 24, enabled event triggers for scheduled tasks on August 25, and announced on September 11 the withdrawal of custom GPTs in favor of plugins.
• ecosystem added Claude Fable 5.1 and GPT 6 Astra to the selector Cowork, switched local control of the Edge browser to general availability, and opened tasks triggered by an email or a Teams message.
• Mistral did not touch Vibe Work during the period; the news is elsewhere (raising of 3 billion euros on September 8, regional inference endpoints in Europe or the United States since August 11).

Where the work is carried out

That's the first question, and it's rarely asked.

At ClaudeA task runs either in the cloud (a "temporary, isolated sandbox on Anthropic's infrastructure," created at startup and destroyed at startup, without access to your network) or on your machine, in a Linux virtual machine isolated from the system by the hypervisor (Apple Virtualization on Mac, Hyper-V on Windows). This local mode, described in the architectural documentationThis is done through the desktop application. It shouldn't be confused with Dispatch, which allows you to send Claude a message from your phone so he can work on the desktop computer (with the application on and open), currently in limited beta on Pro and Max. Local mode provides access to folders you've explicitly shared (since September 4th, you can attach your entire home folder or a whole drive), the integrated browser, and the screen ("Computer use," in beta on Pro and Max, desktop application only, not on Linux). A useful nuance: a cloud session can read and write to your local folders as long as the desktop application remains open; if you close it, the session continues but loses access to the files.

At OpenAIChatGPT Work distinguishes between "Work Cloud" (web, mobile, desktop) and "Work Local" (desktop application only, with access to local files "limited to those required by the task"). This important distinction is explained in the... documentation : "Messages and task context can be stored in the cloud, even when the job is running locally."

At ecosystemCopilot Cowork "Processes files in a temporary, isolated environment within the Microsoft 365 service boundary," and destroys them at the end of the task. FAQ is clear: Cowork "Cannot access or modify files stored locally on your device." It works on OneDrive, SharePoint, Teams, and Outlook, and never deletes files. This is consistent with Microsoft's philosophy: your work is already in the tenant. The exception is local browser use, generally available since August, but only when Cowork is opened in Edge (version 152 and above), and after activation by the administrator.

At MistralVibe Work runs in the cloud, accessible via the web or mobile app, with an integrated editor ("Canvas") for reviewing and editing your work. Work only recognizes the files you upload or those it accesses through a connector. Local disk access exists in Mistral, but only within Vibe Code (the command-line tool, which has "read and write access to the file system and a shell"), not in Work.

What this changes: if your sensitive data is in a local folder and you do not want it to be uploaded, only two platforms give you the choice, and only one (Claude) documents a mode where the agent itself runs on the machine.

How the agent accesses your tools

An agent without tools is a chatbot that writes plans. The four platforms have therefore built catalogs, but on very different scales, and above all with different rules on who has the right to plug in what.

The Claude's directory It displays 824 connectors, including Microsoft 365 (read-only by default, optional write: sending emails, calendar events, OneDrive and SharePoint files, Teams messages), Google Workspace, Slack, Notion, Atlassian, and HubSpot. Salesforce arrived on September 15th as a beta plugin, with 37 pre-written sales skills. Custom MCP connectors are available on all plans, including the free plan (limited to one on the free plan).

ChatGPT OpenAI replaced its App Directory with a Plugin Directory on July 9th. A plugin "can contain skills, apps, or both." OpenAI does not publish totals. A point of note for businesses: on the Business plan, "only administrators and owners can activate developer mode and deploy an app," and "apps cannot be updated after publication" (they must be recreated and republished); the Enterprise and Edu plans have a refresh button. Since September 11th, OpenAI has also announced the gradual phasing out of custom GPTs in favor of plugins.

ecosystem plays a different tune. The official list of plugins It includes four Microsoft plugins (Dynamics 365 Sales, Customer Service, ERP, Fabric IQ) and approximately 145 partner plugins, distributed through the Microsoft 365 App Store, which an administrator must make available. These include Asana, Box, Canva, Gmail, HubSpot, Linear, monday.com, Notion, Trello, Zapier, and Azure DevOps. It does not include Jira, Salesforce, ServiceNow, Slack, or GitHub. Copilot's strength lies in its integration with Microsoft 365 App Store. Cowork is not there: it is native access to your emails, files, meetings, people and conversations, which Microsoft calls Work IQ in its communication and "Enterprise Search" in the list of integrated skills of Cowork.

Mistral The company announces "more than 100 connectors" (Atlassian, Box, GitHub, Gmail, Google Drive, Notion, Outlook, SharePoint, Slack, Stripe, Salesforce, ServiceNow, HubSpot, BigQuery, Snowflake, Databricks, etc.) with full MCP compatibility. Adding a custom MCP connector is "a feature reserved for administrators"; on Free, Pro, and Student plans, the account holder is an administrator by default. Data passing through the connectors is "never used to train or refine" the models.

What this changes: the question is not "how many connectors" but "are mine there, and who has the right to plug them in".

Who approves what?

This is the topic we find most interesting, because it's where trust is at stake. And it's also the one where our version from September 5th has aged the most.

Claude offered three modes per task: "Manually approve", "Automatically approve", "Skip all approvals". Since the merger of September 16The new experience (Pro and Max first) retains only two: "Manual", by default, and "Auto". All three modes remain in place on Team and Enterprise, where the administrator has an "Allow 'Automatically approve' mode" switch and a second one, disabled by default, to allow "Always allow" on connector tools.

ChatGPT There are four levels for apps: "Always ask", "Allow read actions", "Allow low-risk actions", and "Allow all actions". Contrary to what we wrote, the default setting is not "Always ask" but “Allow low-risk actions” Low-risk actions and readings are automatic, while sensitive actions (sending an email or message, creating, modifying, or deleting, changing sharing or access permissions, purchasing, transferring money, or disclosing personal data) require confirmation. In a managed workspace, organizational policy takes precedence over user policy.

ecosystem It requires approval before each important action (Send, Post, or Create button). The options menu is the most granular of the four: for an email or Teams message, you can allow "only to" a recipient, "only to" a domain, or "always"; for other actions, "Approve & don't ask again"; and "Approve All" when multiple approvals are pending. These permissions "only apply to the current session" and can be revoked in the Permissions panel. The interface also indicates the risk level of medium and high-risk actions.

Mistral It requests approval before any action that modifies an external system, with three responses: "Continue" (this time only), "Always allow" (pre-authorizes the function for the session), "Decline". documentation It explicitly recommends "pre-authorizing read-only functions" and "keeping interactive functions requiring manual approval, especially deletions." Permissions are set function by function, for each user; administrators can disable a connector for the entire organization.

What this changes: across the four, the basic reflex is the same. Automatic reading, validated writing. What varies is the precision of the setting, the default value (ChatGPT is the only one to allow "low-risk" writes without prompting), and an administrator's ability to lock down the behavior for the entire organization.

How far does the agent work without you?

All platforms have scheduled tasks. The differences are in the details, and the details speak volumes.

Claude Hourly, daily, weekly, or weekday frequency, on all paid plans. scheduled tasks They run remotely, so at their own pace even when your computer is asleep or the application is closed, and can be attached to a project from which they inherit instructions, files, and connectors. There are no event triggers in Claude; however, they do exist in Claude Code, the developer tool, as "routines" triggered by an API call or a GitHub event, in preview.

ChatGPT 3 active tasks in Free, 5 in Plus, 10 in Business, 15 in Pro and Enterprise, at a maximum hourly frequency on paid plans. Since August 25th, the tasks These actions can be triggered by an event: an incoming Gmail email (filterable by sender or subject), a message in a Slack channel (provided @ChatGPT is added), or activity on a GitHub pull request, with a limit of 30 executions per hour and 720 per day. On Enterprise and Edu, administrator authorization is required. Workspace Agents, a separate feature available only on Business and Enterprise, go further: an agent with its own memory, skills, and a Slack channel where it responds when mentioned.

ecosystem 25 scheduled prompts per user, and since August, tasks triggered "when you receive an email," "when a message arrives in a Teams channel," or "in a Teams conversation," with the option to only react to @mentions. These tasks are in "write then get approved" mode by default, with bandwidth limits and loop protection. The "goal" skill that we spotted in the interface in early September doesn't appear anywhere in the documentation; therefore, we no longer mention it.

Mistral 5 tasks on the free plan, unlimited on the Pro and Team plans (within reasonable usage limits). You choose a single frequency: daily, weekly, monthly, or annually, and a specific time; no hourly execution. No event-based triggers either: "triggers are time-based for now; event-based triggers may follow."

What this changes: two platforms (ChatGPT, Copilot) have taken the step of using agents that react to events rather than clocks. It's a change in nature, not degree: the agent becomes a standby colleague.

How do we pass on know-how to him/her?

This is the clearest point of convergence. All four platforms have adopted "skills": a folder containing a SKILL.md file with instructions, a checklist, and an output template. Mistral explicitly states this: it's the open standard "Agent Skills," "initially developed by Anthropic and now adopted by several agent clients." OpenAI refers to the same standard. Microsoft accepts the direct import of a .zip or .skill archive containing a SKILL.md file in the root directory.

Each vendor adds its own layer. Claude and ChatGPT package skills and connectors into shareable plugins (with Claude, a plugin "groups skills, connectors, and sub-agents"; with ChatGPT, administrators can synchronize a plugin marketplace from GitHub). Microsoft caps skills at 50 per user, stores them in OneDrive, and rates each skill from 0 to 100 based on four criteria (clarity of trigger, precision of instructions, scope, robustness). Mistral delivers twelve built-in skills, allows administrators to customize them, and complements them with Libraries, persistent document repositories that can be shared with Collaborator and Reader roles (15 GB in Pro, 30 GB per user in Team).

What this changes: know-how documented in a SKILL.md file becomes portable from one platform to another. This is good news for teams that don't want to rewrite their procedures every time they switch tools.

The model and the effort

A small interface detail that says a lot about each person's philosophy.

Claude It displays the model and effort level separately, next to the submit button, with five notches (Low, Medium, High, Extra High, Max). Since September 1st, the most advanced model is Fable 5.1, and tasks start by default at Medium effort. documentation simply warns that a higher effort "consumes more tokens, so you will reach your limits faster".

ChatGPT All of this has been replaced by a "Power" slider, ranging from "Faster" to "Smarter," which adjusts both the model and the reasoning. On eligible accounts (Pro, Business Premium, Enterprise), it has six positions since the arrival of GPT-6 Astra: Terra Light, Sol Light, Sol Medium, Astra Light, Astra Medium, and Astra Extra High. "The options may differ depending on the plan and deployment stage." The user no longer chooses a model; they choose a power level.

ecosystem It offers "Auto" as the default setting and allows a choice between seven models from the two families: GPT 5.5, GPT 6 Sol, GPT 5.6 Terra, GPT 6 Astra, Opus 5, Claude Sonnet 5 and, since September, Claude Fable 5.1. level of reasoning effort ranges from Light to Max (Medium by default), with an honest note: "a higher effort is more thorough, but slower, and consumes more credits."

Mistral It offers neither a choice of templates nor a switch in Work mode: "step-by-step reasoning starts automatically when the task requires it." The Fast/Think selector only remains in the old Chat mode. Vibe runs on Mistral Medium 3.5 by default.

What it really costs

The prices listed don't tell the whole story, because agent work is much more time-consuming than a conversation. And three of the four publishers have revised their pricing schedules since the end of August.

At ClaudeTasks are included in the paid plans: Pro at $20 per month ($17 annually), Max starting at $100 (5 or 20 times the Pro usage), Team at $25 per seat ($20 annually) with a Premium seat at $125 ($100 annually), and Enterprise at the seat price plus API usage. A 5-hour rolling window and a weekly cap apply to all products, including the desktop application and Claude Code. "Multi-step tasks consume more resources than a quick question," it states clearly.

At OpenAIChatGPT Work "follows the same usage structure as Codex." Each seat includes usage, measured over a 5-hour window for Standard seats; beyond that, "eligible usage can continue with credits purchased by the workspace, if spending controls allow it," otherwise, "features that require credits may be unavailable until credits are added." The Business tier was split on August 24: Standard at $25 per user ($20 annually), Premium at $125 ($100 annually) with five times more usage and no 5-hour limit. For individuals: Plus at $20, Pro at $100 or $200, the latter closed to new subscriptions since September 10. The free plan only has a limited version of Work on the desktop application.

At ecosystem, Cowork is billed on a metered basis, starting at $0,01 per credit, in addition to the Microsoft 365 Copilot license at €26 excluding VAT per user per month (Enterprise version). The following consume credits: model responses, tool and skill calls, image generation, and browser tasks. The administrator must have enabled metered billing and created a spending policy that includes Cowork ; The documentation It clarifies that "a spending policy is about access control, not just a budget." Starting November 2nd, new Microsoft 365 Copilot Business licenses sold through CSP partners will include pay-as-you-go billing by default.

At MistralThe Work features, connectors, and scheduled tasks are available even in the free plan, with some limitations, then $14,99 for Pro, $24,99 per user for Team ($50 per month minimum), and $5,99 for students. Workflows are reserved for paid plans.

Where does your data go?

We conclude with the topic that should open any business discussion.

Mistral Vibe hosts in the European Union by default. But be aware: on Vibe Standard (Free and Pro), users are not opted out by default for training; you must manually disable "Allow the use of your interactions to train our models" in the privacy settings. In Teams, administrators decide. Only Enterprise is automatically opted out.

Claude Claude.ai stores data in the United States by default; no European residency is documented for claude.ai, only inference via Bedrock, Vertex, or Foundry in the European region is supported, for businesses. The Free, Pro, and Max plans have a "Help improve our AI models" setting (up to 5 years of retention if enabled, deletion within 30 days otherwise). Team, Enterprise, and API plans do not train on your data by default.

ChatGPT does not, by default, affect data in the Business, Enterprise, Edu, and Healthcare workspaces. data residence In Europe (European Economic Area and Switzerland) is reserved for Enterprise, Edu, Healthcare and the API; the Business pricing page indicates "data residence: not available".

ecosystem applies the EU Data Boundary to Copilot, with two documented exceptions: web queries sent to Bing, and Anthropic models, "currently excluded from the EU Data Boundary." For European and UK tenants, these models are disabled by default and require a activation by the administratorChoose Opus 5 or Fable 5.1 in CoworkTherefore, for a European company, it is a compliance decision before it is a performance decision.

What this changes: "hosted in Europe" and "not used for training" are two separate questions, and none of the four platforms answers yes to both on all counts.

The comparison at a glance

	Claude	ChatGPT Work	Copilot Cowork	Mistral Vibe Work
Place of performance	Cloud, or Linux VM on the machine (desktop application)	Cloud, or local on the desktop app	Isolated environment in tenant M365	Cloud
Local files	Shared folders or drives, screen (Pro, Max)	Desktop app, task files	No (OneDrive, SharePoint)	No (upload)
Tool catalog	824 connectors, MCP on all planes	Plugin Directory, MCP custom by admin	Native Work IQ + around 150 plugins	100+ connectors, custom MCP by admin
Approvals	Manual / Auto (3 modes in Team, Enterprise)	"Allow low-risk actions" by default	By recipient, domain or session	Continue / Always allow / Decline, by function
Scheduled Tasks	Minimum time, no event	10 in Business, schedule, events, Gmail, Slack, GitHub	25, email and Teams events	5 in Free, minimum daily
Choice of model	Model + 5 levels of effort	6-position "Power" slider	Automatic, 7 GPT or Claude models + 5 effort levels	None, automatic reasoning
Cost	Included: 5-hour window + weekly ceiling	Usage included, then credits; Standard seats $25 or Premium seats $125	License €26 excluding VAT per user per month (enterprise version) + credits at $0,01	Starting with the free plan, $14,99 for Pro
Data	US storage; no team training	No training in Business; EU residence in Enterprise	EU Data Boundary, except for Bing and Anthropic models	Hosted in the EU; opt-out training to be activated in Free and Pro versions

What you must remember

Agent mode is no longer a promise, it's a tab, and at Claude's it's not even a tab anymore. Choosing a "cowork" comes down to answering three questions in order: where does my data live and where does the agent have the right to go, who validates the actions that modify something, and to what extent do I accept that it works without me, from a one-off task to permanent monitoring.

The four vendors have responses consistent with their history. Microsoft starts with your tenant, Mistral with European sovereignty, ChatGPT with computing power measured in credits, and Claude with the agent that works on your files, whether in the cloud or on your machine. None of them are neutral. And the first version of this article, two weeks old, already contained four outdated points: that's the pace we should expect.

Methodology: The September 5th version was based on a live walkthrough of the four interfaces. This September 20th update was created exclusively from the official documentation (support.claude.com, help.openai.com and learn.chatgpt.com, learn.microsoft.com, docs.mistral.ai and help.mistral.ai), retaining only pages dated August or September 2026. When a point observed in early September in the interface did not appear in any documentation, we removed it rather than keeping it.