Microsoft Copilot Cowork's New Pricing: What It Actually Costs, and How to Budget for It

Microsoft has changed how you pay for one of its most powerful AI tools, and a lot of businesses are about to get a surprise on their next invoice. On 16 June 2026, Copilot Cowork became generally available worldwide, and with it came a pricing model that works very differently from the flat monthly fee most people expect from Microsoft 365.

I have spent the past few weeks running real tasks inside Cowork and building a budget model from what I saw, because I make a rule of testing a tool properly before I tell anyone how it works. What I found is that the new pricing is genuinely useful for some organisations and genuinely risky for others, and the difference comes down to whether you understand it before you switch it on. Here is the plain-English version for anyone who has to sign off the bill.

What actually changed

Until now, Copilot has been priced the way most Microsoft products are: a predictable per-user, per-month fee. You knew exactly what each person cost you every month, regardless of how much they used it.

Copilot Cowork breaks that pattern. It is Microsoft's new agentic tool, the part of Copilot that runs complex, long, multi-step tasks on its own and hands you back a finished result rather than a draft. Think aggregating a dozen reports into a financial model, or working through a stalled sales pipeline and returning a ranked list of what to chase. That kind of work uses far more computing power than a quick chat prompt, and Microsoft has decided to charge for it accordingly.

So Cowork now uses two layers of cost. First, every user still needs a Microsoft 365 Copilot licence, the same per-seat fee that powers Copilot in Word, Excel, Teams and Outlook. Second, on top of that licence, you pay for the actual Cowork tasks people run, billed by usage. That second layer is the new and unpredictable part.

How the new billing works: Copilot Credits

The usage layer is measured in something Microsoft calls Copilot Credits. One credit costs $0.01 under pay-as-you-go pricing (these are US list prices, so expect the UK figure to track the dollar rate plus VAT).

The number of credits a task burns comes down to four things:

• Model use: which AI model runs the task, and how hard it has to reason.
• Context retrieval: how much of your own data it pulls in for grounding.
• Tool calls: how many actions it takes across your apps.
• Runtime: how long it runs from start to finish.

A quick, single-step job is cheap. A long, multi-source, multi-output project costs a lot more. That is the entire logic of the model: you pay in proportion to the work done.

One detail worth flagging for finance teams: Copilot Credits are a shared Microsoft currency. The same credits can be spent across other usage-based Microsoft AI services, not only Cowork, so they need to be tracked as a pooled cost rather than a single line item.

Light, medium and heavy: the only mental model you need

Because the exact credit cost of any single task moves around, Microsoft groups work into three tiers, which is the most useful way to think about it:

• Light: a few sources, limited reasoning, one output or fewer. For example, "sort my inbox from the last 24 hours."
• Medium: several sources, structured reasoning, two or more outputs.
• Heavy: broad aggregation, deep reasoning, many outputs. For example, "combine these twelve reports, build a financial model and produce a deck plus a summary memo."

The practical takeaway is simple. The cost of Cowork in your organisation is driven almost entirely by how many heavy tasks your people run, and how often. A team that mostly does light tidying-up work will see a small bill. A team that leans on Cowork for big analytical jobs all day will see a large one.

How to estimate your spend before you commit

The question I heard most while testing this, and the one Microsoft itself says it hears most from customers, is the obvious one: how on earth do you budget for a variable bill?

The answer that works is to estimate by persona, not by task. You do not need to predict every job your team will run. You group your people into a handful of usage types, for example senior leaders, customer-facing staff, technical staff and general knowledge workers, give each group a rough monthly mix of light, medium and heavy tasks, multiply by the credit price, and add it up. Microsoft observed four common personas during its preview and, in its general availability announcement, offers a free Customer Cowork Estimator spreadsheet that does the maths for you, a sensible starting point even if you refine the numbers later.

This is exactly the kind of model I built while testing Cowork. To make it concrete, I put together a simple budget that estimates the token cost of typical workplace tasks, an inbox catch-up, a slide deck, a research report, and rolls them up into a monthly figure per user. Here is a snapshot, using a frontier-model rate for illustration:

The headline lesson from building it was clear: cost is driven far more by the high-frequency light tasks and the occasional heavy one than by anything in the middle, and the first estimate is always rough. The only way to get a real number is to turn Cowork on for a small group, watch the usage reports for a few weeks, and then scale. Treat your first month as paid research, not as your steady-state cost.

The bit everyone is worried about: bill shock

It would be dishonest to write about this pricing without naming the elephant in the room. Usage-based AI pricing has a track record of nasty surprises.

The clearest warning sign came from Microsoft's own stable. GitHub Copilot, the developer version, moved to usage-based billing at the start of June 2026, and some developers watched their monthly costs jump from around $29 to several hundred dollars almost overnight. These were technical users who understood exactly what they were doing, and they were still caught out when the billable unit changed underneath workflows they had already got used to. The honest read is that if engineers can be surprised by credit burn, ordinary business users with no visibility into what a task costs are even more exposed.

To be fair to Microsoft, the model is not unreasonable in principle. Charging for heavy compute by usage means light users stop subsidising heavy ones, and for an organisation that uses Cowork sparingly it can genuinely work out cheaper than a flat add-on fee, because you only pay when the tool actually does work. The risk is not the model itself. The risk is switching it on without guardrails and finding out what it costs the hard way.

The good news: the controls are real, and it is off by default

This is the most reassuring part, and the part worth telling your IT team. At launch, Copilot Cowork is switched off by default. Nobody in your organisation can run a single billable task until an administrator deliberately turns it on. There is no accidental spending.

Beyond that, Microsoft has built genuine cost controls around three ideas:

• Control: admins decide who gets access and can set hard spending caps at the level of the whole organisation, a group, or an individual user.
• Visibility: usage is reported by user, group and feature, so you can see exactly where credits are going, and a per-task price display for users is on the way.
• Efficiency: choose pay-as-you-go for flexibility or commit to a prepaid volume for a discount, and where more than one model is available, route routine work to a cheaper model and keep the expensive one for hard problems.

Used properly, this turns a scary open-ended bill into something you can govern like any other budget. The mistake to avoid is enabling Cowork broadly on day one. Enable it for a small pilot group with a firm cap, learn your real costs, then expand.

How it compares to Claude Cowork

Copilot is not the only game in town, and the most direct comparison for the kind of agentic work Cowork does is Anthropic's Claude Cowork. I use Claude Cowork in my own work, so this is a comparison I can speak to from both sides, and it is worth understanding because the two are priced on the same basic shape but with very different details.

The important thing is that the two are priced on different philosophies. Copilot Cowork is pay-per-use: you pay the Microsoft 365 Copilot licence, roughly $30 per user per month on top of a qualifying base plan, and then every Cowork task is metered in Copilot Credits at $0.01 each. Your bill scales with use, and it can spike.

Claude Cowork, for most people, is a flat monthly subscription. Pro is $20 a month, and Max is $100 or $200 a month for heavier use, and each one includes Claude chat, Claude Code and Cowork with generous usage limits and no per-task billing. There is also a Team plan at around $25 per seat, and an Enterprise plan at $20 per seat plus metered usage for large organisations that want the same consumption model Microsoft uses. So the real contrast for a budget holder is this: with Copilot you pay for exactly what you consume and accept some uncertainty, while with Claude's Pro and Max plans you pay a fixed fee and work within usage limits, which caps your cost but does not flex down if you barely use it.

The differences that matter for a decision are these. Copilot Cowork is built inside the Microsoft 365 trust boundary, so if your business already runs on Microsoft, your data governance, audit and compliance controls extend to it automatically, which is a genuine advantage for regulated organisations. Claude Cowork is the stronger pick for teams that want Claude specifically, want coding bundled in, or work across a mixed software stack rather than an all-Microsoft one. Microsoft has published its own benchmark claiming Copilot Cowork came out 30 to 40 percent cheaper per prompt than Claude Cowork with the Microsoft 365 connector, but that is a vendor test rather than an independent one, so I would treat it as directional, not gospel. And worth noting for timing: Anthropic is currently offering $1,000 of credits per Enterprise seat activated by 2 July 2026, which mirrors Microsoft's own end-of-June deadline almost exactly. Both vendors are clearly racing to lock in seats before the other does.

Put the two philosophies side by side per user and the trade-off is clear. Copilot's figures are illustrative estimates that move with your task mix, while Claude's are the actual flat plan prices:

The honest summary is that there is no universally cheaper option. The cheaper one is whichever fits your existing stack and your actual task mix, which is exactly why modelling your own usage matters more than any vendor's headline number.

The deadline that matters right now

If your organisation took part in Microsoft's Frontier preview programme between 30 March and 16 June 2026, there is a clock ticking. Those tenants get a grace period during which Cowork usage is not billed, but that grace period ends and full billing begins on 1 July 2026. Customers who used Cowork through Frontier need to enable billing before then to retain access to the tool.

In other words, this is a decision you cannot quietly ignore. Either you enable billing, with sensible caps in place first, or you lose the tool. If you were in the preview, the single most useful thing you can do this week is confirm the exact cutover date with whoever manages your Microsoft tenant, so the first invoice is not a shock.

What I would do if it were my business

Having actually run the tasks and modelled the costs, here is the short version of my advice for any leader weighing this up.

1. Do not switch it on for everyone at once. Start with a small group who will get the most value from heavy, time-saving tasks, and set a hard spending cap before they touch it.
2. Treat month one as research. Watch the usage reports closely and use them to learn your real costs, rather than assuming a settled budget.
3. Separate heavy tasks from quick prompts. Decide deliberately which work is genuinely worth a heavy task and which is just a chat prompt that does not need Cowork. That single distinction drives most of your bill.
4. Watch the model options. Microsoft has a cheaper model coming specifically for routine work, which will change the maths again.

Usage-based pricing rewards organisations that are deliberate and punishes those that are not. The tool is powerful, and for the right tasks it genuinely collapses a week of work into a morning. But the businesses that come out ahead will be the ones that understood the bill before they ran up the first one.

If your team is rolling out Copilot or Cowork and wants this turned into a confident, governed adoption rather than a budget surprise, that is exactly the kind of thing I help with through Microsoft Copilot training and hands-on AI sessions for non-technical teams.