For two years, the conversation about AI at work was about licenses. Thirty dollars per user per month, approved or not approved, done. A fixed number finance understands and that fits in a spreadsheet.
Copilot Cowork broke that comfort. It's Microsoft's agent that doesn't hand you a draft but the finished work: it goes through your email, cross-references meetings, builds the project board, writes the report with citations. And it charges for each of those things, on top of the license.
In September, Microsoft added the missing piece: users now see, on the same screen, how much the task they just asked for cost. The agent came with a taxi meter. And the question stopped being "how much does Copilot cost?" and became "how much does this Tuesday cost?".
The agent born with a borrowed brain
A bit of history before we talk money. Microsoft introduced Copilot Cowork on March 9, 2026, integrating technology from Anthropic's Claude family, the same company that was already selling its own Claude Cowork. It reached worldwide general availability on June 16, after a preview that more than half of the Fortune 500 went through.
Today it runs on Anthropic's Opus 4.8 and Sonnet 4.6, OpenAI's GPT 5.5 and, since September 4, GPT-6 Astra. Microsoft also announced Cowork 1, an in-house model tuned to cost less.
The practical difference from the original Claude Cowork is territory: Anthropic's works with local files and applications, while Microsoft's lives in the cloud, inside the Microsoft 365 tenant, with auditing, eDiscovery and inherited sensitivity labels.
Hold on to that borrowed brain detail. It comes back later.
The bill has two floors
Copilot Cowork is charged in two layers. The first is the Microsoft 365 Copilot license, which is a prerequisite: 30 dollars per user per month for large enterprises, 20 for the Business version. Without it, Cowork doesn't exist.
The second is consumption, measured in Copilot Credits. On pay-as-you-go, each credit costs one US cent. Those who commit to volume upfront (the P3 plan) get a discount.
Four things consume credits: the model running the task, the context the agent pulls from your email, files and meetings, the tools it uses (sending an email, scheduling, editing a document) and runtime. Tasks with many sources, deep reasoning or multiple deliverables cost more.
Microsoft published reference ranges:
| Task type | Approximate credits | Approximate cost | Example |
|---|---|---|---|
| Light | 100 to 300 | 1 to 3 dollars | Reviewing the calendar, triaging messages |
| Medium | 400 to 700 | 4 to 7 dollars | Building a project board from email |
| Heavy | over 700 | over 7 dollars | Research brief with citations |
Now let's do the math the brochure doesn't. Someone who delegates two medium tasks per business day spends, over a twenty-day month, between 160 and 280 dollars in credits. On top of the 30 for the license. Multiply that by a team of ten, and the "30-dollar assistant" becomes a two to three thousand dollar monthly budget line.
The license gives you the car keys. Credits are the fuel. And nobody told you the mileage until now.
September: the meter becomes visible
Until recently, the real cost of a task was a mystery discovered in the admin report. In September, that changed.
Users now see their session cost, credits used, the percentage of monthly quota left and the date it resets. A button also shows group spend, because spending policies are shared among colleagues. And there's a command to check credits consumed in the middle of a task.
On top of that came a reasoning effort slider. The interface itself puts it bluntly: more effort is more thorough, slower and uses more credits. It's the first time a Microsoft 365 product puts a cost decision right in front of the everyday user, task by task.
On the admin side, the tools were already there: Cowork is off by default in every tenant, and spending caps can be set per tenant, group and user, with alerts before hitting the limit.
There's an interesting nuance in how Microsoft measures value. Its consumption dashboard reports "assisted hours of work" across eight task types, and the company itself describes it as a deliberately conservative proxy, not a finance number. Translation: they show you the exact cost and the estimated benefit. One of those is accounting. The other is cautious marketing.
The benchmark defending it was run by the defendant
Here's the part that deserves a slow read. In the general availability announcement, Microsoft published a comparison: across 125 runs with light, medium and heavy prompts, Copilot Cowork came out 30 to 40% cheaper than Claude Cowork using its Microsoft 365 connector. Both on the same model, Opus 4.8.
Sounds decisive. But there are three details.
The one measuring is the one selling. It's not an independent lab, it's the vendor comparing itself to a competitor. We already covered here how even the industry's most respected rankings get distorted when money is on the line. An internal study, however serious, plays at home.
Claude's cost was calculated at public API rates. Microsoft estimated the Claude Cowork side by adding up tokens at list price plus connector usage. It's a valid way to measure, but it isn't necessarily how someone already using Claude on a subscription plan pays. Comparing your volume-discounted price against the other side's list price is a classic of every sales deck.
The brain is the same. Both columns of the comparison run on Opus 4.8, an Anthropic model. What Microsoft is really saying is that it orchestrates the same engine better inside its own ecosystem. That may be true: it has your data at home and doesn't need to fetch it through a connector. But that's an advantage of location, not of intelligence.
None of this means the number is false. It means it's a starting point for your own measurement, not a conclusion.
How to avoid a nasty surprise on the bill
If your organization is evaluating Cowork, a few practical rules:
- Start with a pilot group and a low cap. Group limits exist for exactly this. A month of real usage is worth more than any reference table.
- Measure your tasks, not Microsoft's. The cost command and the user panel show what each task spent. Write down the recurring ones and calculate the real monthly cost per person.
- Teach the effort slider. Most tasks don't need the maximum. A team that leaves everything on high effort is paying for reasoning it doesn't use.
- Compare against human hours, not against zero. Seven dollars for a cited report is expensive if nobody needed it and very cheap if it replaces half a day of an analyst's work.
- Look at total cost before comparing vendors. License plus credits on one side, subscription or API on the other. The fair comparison is against your usage, not someone else's benchmark average.
And if you're still not sure which Copilot tier your company has today, start here.
The point
Copilot Cowork marks a turning point in how AI at work gets paid for. The flat license was the model for the assistant that suggests. Per-credit consumption is the model for the agent that executes. And an agent that executes, like any colleague, costs according to how much it works.
That Microsoft now shows the meter is good news. That the study saying its taxi is cheaper was run by Microsoft itself is a reminder that the final word on cost belongs to your invoice, not to the sales deck.
If you want to understand what AI really costs before the bill arrives, subscribe to the macareno.net newsletter. Every week, a little less hype and a little more context.
Sources
- Copilot Cowork is now generally available, AI at Work Blog (Microsoft)
- Microsoft launches Copilot Cowork with usage-based pricing, Computerworld
- Microsoft's new Copilot Cowork integrates Anthropic's Claude, GeekWire
- Copilot Cowork Pricing 2026: How Usage-Based Billing Works, Quisitive
- Microsoft Copilot Cowork: New 2026 Pricing Guide, A Guide to Cloud & AI
- What's New in Microsoft 365 Copilot: September 2026, A Guide to Cloud & AI
- NEW Copilot Updates Are Here: September 2026, M365 Copilot Connection