Microsoft prices by task complexity, not by tokens. Here is how to turn the published light, medium, and heavy ranges into a defensible monthly budget by persona, before you provision a single seat.
Copilot Credits bill by how hard a task is, not by what you sent. That makes the meter feel unpredictable, but it is straightforward to model. Estimate the task mix per persona, apply the published credit ranges, and convert at one cent per credit. This guide builds that model, shows why the heavy tier decides your budget, and how to validate it after go live.
Microsoft publishes illustrative ranges by intensity, denominated in credits. The dollar figure is just the credit count times one cent. The ranges are set out on the Copilot Studio pricing page and the Copilot Credits overview.
Cowork task tiers in credits and dollars
| Intensity | Example | Credits | Cost at one cent |
|---|---|---|---|
| Light | Draft a short update from calendar and notes | 70 to 200 | $0.70 to $2.00 |
| Medium | Meeting brief from emails, CRM, and a deck | 400 to 600 | $4.00 to $6.00 |
| Heavy | Analyze a large data export, write a summary | Over 1,500 | Over $15.00 |
Because intensity tracks how much the agent reads and writes. A token is about three quarters of a word, so a light task reads a short stack of context and a heavy task can read hundreds of pages before it writes anything.
White Paper · Microsoft
What Microsoft Copilot Cowork Really Costs
What a task really costs in dollars, the prepay floor, and the same work on Claude direct. Read it free.
You cannot convert a credit cleanly into a token rate, because the credit bundles more than the model. But the rough token shape behind each tier helps sanity check whether a workload will land light, medium, or heavy.
The lesson is that context size, not the prompt, decides the tier. A short instruction that pulls a giant data export is a heavy task. Designing agents to pull only the context they need is the cheapest optimization you have.
The model is a grid: personas down the side, task tiers across the top, volume in the cells. It takes an hour and it is the most useful number you can hand finance.
Your Copilot budget is not set by your user count. It is set by how many heavy tasks those users run.
Take 200 users, each running about 20 light, 10 medium, and 2 heavy tasks a month. That is the deployment the pillar prices, and it lands near 27 million credits a year, roughly 271,000 dollars at list.
Monthly draw per user, by tier
| Tier | Tasks per user | Credits each | Share of the bill |
|---|---|---|---|
| Light | 20 | ~135 | Small |
| Medium | 10 | ~500 | Moderate |
| Heavy | 2 | 1,500 plus | Largest |
The full four way comparison against Claude direct is in the credits versus Claude direct guide, and the meter in the Copilot Credits pillar.
The model is a planning band, not a promise. Once you are live on pay as you go, the admin center gives you the real distribution, and you refine.
Three, and all of them understate the budget.
Build the model before you provision.
Microsoft prices by task complexity, not by tokens, and publishes illustrative credit ranges. A light Cowork task is 70 to 200 credits, a medium task 400 to 600, and a heavy task over 1,500. At one cent per credit, that is roughly 70 cents to over 15 dollars per task.
How much the agent reads and writes. A heavy task ingests a large context, sometimes hundreds of pages of mail, files, and data, then writes a full analysis. Input tokens are everything the model reads, output tokens are what it writes back, and both push the credit count up.
Estimate light, medium, and heavy task volume per persona per month, multiply each by its credit midpoint, sum across personas, and convert at one cent per credit. Model the high end of the heavy range, because heavy tasks dominate the total even at low volume.
Because the credit spread is more than seven to one between light and heavy. A handful of heavy analyses per user per month can outweigh dozens of light drafts. The forecast is far more sensitive to the heavy task count than to the light one.
Yes, as a separate line. Work IQ queries run 20 to 150 credits and tool actions a tenth of a credit each, drawn from the same pool. Custom agents that run many grounded queries can exceed your human users, so forecast them per agent, not per person.
They are illustrative, not a quote. Real consumption varies with how much context each task pulls and how the agent is configured. Treat the ranges as a planning band, run pay as you go to get your real distribution, and refine the model against actuals after a quarter.
The task mix model in dollars, the credit to dollar conversion across light, medium, and heavy work, the build versus buy math against Claude direct, and the governance controls to set before you provision.
Used across more than five hundred enterprise engagements. Independent. Buyer side. Built for procurement leaders sizing Copilot inside the next EA renewal.
500+ enterprise clients. 11 vendor practices. We model your task mix in dollars, weigh direct Claude against buying through Microsoft, and set the governance and commitment terms before Microsoft sizes them for you.
One email a month on Copilot Credits, EA renewal levers, Azure commitment, and audit defense. Buyer side only.
