Claude Cowork vs Copilot Cowork: what each one costs, and the open-source option you own

Claude Cowork is the polished agent a single desk can delegate to, and Copilot Cowork runs that same job inside Microsoft 365, and Kortix is the open-source CompanyOS whose agent's definition, grants, and record of work are files in a repository the company owns. Claude Cowork charges for a seat and caps its use; Copilot Cowork requires a license and meters the work in Copilot Credits above it. Microsoft publishes a credit rate, per-task estimates, and an estimator for sizing its meter. Anthropic publishes a seat price and a limit. Neither leaves you a file that defines the agent.

Two products built on the same bet

Anthropic built Claude Cowork for knowledge work. Hand it a goal and it works inside whichever folders and tools you point it at. It returns a memo, a deck, or a spreadsheet to review. Anthropic's release notes put the change on September 16, 2026, the day Cowork opened from any conversation. Pro and Max subscribers got it first, on web, desktop, and mobile.

Copilot Cowork runs that job inside Microsoft 365. Microsoft's Cowork guidance describes long-running, multi-step workflows across apps, files, and data, grounded in the tenant's mail, meetings, and files through Work IQ. It needs a Microsoft Copilot license. IT administrators control who can access it. The license does not cover the work. Cowork bills through usage-based billing in Copilot Credits, alongside Copilot in SharePoint and the Work IQ API (billing overview).

What a credit buys, and what a seat covers

Copilot Credits are one currency across Microsoft's metered AI services, and Cowork draws on the same meter. On Microsoft's pay-as-you-go Copilot Studio meter a credit costs $0.01, the figure in the Copilot Credits Guide for October 2026. The Copilot Credit Pre-Purchase Plan, a year paid up front, discounts that rate by 5% to 20%. Both rates belong to the shared meter; Cowork has no price list of its own.

Microsoft does publish planning inputs: the guide estimates consumption for representative task types. It labels the figures rough planning estimates and points to a Copilot Credit Estimator for a monthly model. Its method for Cowork groups people by persona, estimates the prompts each one runs, splits them into light, medium, and heavy work, then applies an average price per prompt. The prompt count is the input only you can supply. Cowork also picks the model per task, trading performance against cost, so two tasks of the same shape can price differently (Cowork guidance).

Anthropic sells a seat instead. Claude's pricing page puts Pro at $17 a month on the annual plan, or $20 month to month; the annual plan is paid as $200 up front. Max begins at $100. A standard Team seat is $20 on the annual plan and $25 month to month. Enterprise runs $20 per seat with usage billed at API rates (checked October 2026). Cowork ships with Pro, Max, Ultra, and Team; Enterprise accounts opt in (Cowork page). The seat carries one shared pool. A session limit resets every five hours, a weekly limit every week, and every Claude surface draws on the same allowance (usage limits). Anthropic publishes no figure for the tasks a seat covers; a buyer sets that cap from experience.

Where the definition and the record live

An administrator shapes the agent in both products, and the shape stays in the console. On the Teams and Enterprise tiers an admin decides who gets in, caps what each group may spend, and scopes tool permissions by department. OpenTelemetry carries the activity out to a SIEM (Cowork page). Microsoft's billing guidance keeps spending policies, limits for the organization or one person, alerts, and credit requests inside its 365 admin center. Consumption is reviewed there by spending policy, person, group, agent, and service.

A Microsoft admin reads credit usage per person and per service; a person who hits a limit requests more. On Claude Enterprise, usage analytics sit beside the admin controls, with an Analytics API and OpenTelemetry for observability. Both hand a security team a console or a feed it already reads. Neither keeps the agent's definition in a repository the customer owns.

Side by side

Where each one fits

Copilot Cowork operates inside Microsoft 365, where the business already runs and the licenses are in hand. Work IQ grounds the work in the mail, meetings, and files inside the tenant boundary. The credits join a contract the company already has, and one admin center holds access, limits, and the usage record.

Claude Cowork fits one person who wants a strong agent across web, desktop, and mobile: a memo researched overnight, a folder tidied up. At $17 a month on the annual plan, the seat is small enough to expense without a procurement cycle. When a job needs a desktop app, Cowork either drives its own browser in a side panel or takes the screen over. Anthropic's page says it reaches for connectors and integrations first.

End the contract and the agent stays where it was set up, in Claude's product or Microsoft's 365 admin center. A company that would rather review its operating rules the way it reviews code gets that from Kortix. A member's connected computer gives agents its files, shell, and screen under the rules set there (Computers).

The same settings, kept in your repository

Every Kortix project lives in a git repo, and the settings a vendor would keep in a console sit in its manifest instead. The starter divides the work in two: the agents: block declares governance, and the markdown file it points at carries the behavior.

agents:
  kortix:
    prompt_file: agents/kortix.md
    connectors: all
    secrets: all
    kortix_permissions: all
    skills: all

That prompt_file: key is what a console cannot give you. Because the agent lives at agents/kortix.md in the repo, any change to its behavior shows up as a diff on a change request. The four grants next to it decide how far it reaches: the connectors it can call, the secrets it can receive, the Kortix permissions it holds, the skills it loads. A comment in the manifest notes that a grant left out resolves to none, which keeps a specialist agent locked down at the start. The starter hands its default agent all four; a team trims them once it knows what the project needs. Scheduled work is another line in that file, reviewed next to the agent that runs it.

triggers:
  - slug: weekly-report
    type: cron
    agent: kortix
    cron: "0 0 15 * * 0"
    prompt: Build the weekly report.

Session work reaches main through a change request. Merge is default-deny for agents, and the permission that overrides that is project.gitops.merge (Kortix documentation). Whoever holds it reads the diff, including one that widens a grant, and decides whether it lands. The starter's kortix_permissions: all carries that capability too, so a team that would rather have a human approve every merge removes it from the list.

Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. Any model provider you like works with keys you hold. The free tier costs $0 and gives you 500 credits each month on your first workspace, drawing on keys you supply or a ChatGPT subscription. Paid credit packs begin at $50 for 5,000 credits, which is 100 credits a dollar (Kortix pricing).

Three other posts work through the terms. Kortix vs Claude Cowork walks the closed product line by line, and the Copilot pricing post works the credit arithmetic. The alternatives post compares Cowork-shaped projects.

Start on Kortix Cloud at no cost, or self-host it for free: Try Kortix.