Copilot Cowork Cost Per Task: Light, Medium & Heavy

The Copilot Cowork cost per task is the most useful number you can carry into a budget meeting, because Cowork bills by the task and tasks are not equal. Microsoft sorts them into three bands — light, medium, and heavy — that span a seven-to-ten-times cost range. Knowing what each band runs, and what pushes a task from one to the next, is how you turn a variable meter into a forecast.

Every Cowork task is priced in Copilot Credits at $0.01 each, derived from four inputs: model use, context retrieval, tool calls, and runtime. We unpack those inputs in the Copilot Credits explainer; here the focus is the output of that calculation — the per-task cost — and how to plan against it. For the full bill including the seat floor underneath, see the Microsoft Copilot Cowork cost guide.

The three task bands, defined

Microsoft observed three task patterns during the Frontier preview and built its cost guidance around them. The definitions are behavioral, not arbitrary:

• Light tasks use a small number of knowledge sources, apply limited reasoning, and produce one output or fewer. Think summarizing a document, drafting a short response from a single thread, or extracting a few fields. These run roughly 100-300 credits, or $1-$3.
• Medium tasks draw on multiple sources, apply structured reasoning, and generate two or more outputs. A competitive briefing or a multi-document synthesis sits here, in the rough range of 300-700 credits, or about $3-$7.
• Heavy tasks aggregate broadly, apply deep reasoning, and produce many outputs. Comparing thousands of files or triaging an entire pipeline lands here, at 700+ credits, or $7 and up.

What each band looks like in practice

Microsoft's own launch examples map cleanly onto the bands. A team had Cowork compare nearly four thousand files across two product versions — broad aggregation, deep reasoning, many outputs, unmistakably heavy, comfortably north of 700 credits. A sales lead pointed Cowork at a stalled pipeline and got back a ranked list of at-risk opportunities with the specific follow-up that had gone cold on each — multiple sources and structured reasoning, a solid medium-to-heavy task depending on pipeline size. A knowledge worker summarizing a single report sits at the light end.

The lesson is not the absolute numbers; it is the spread. A heavy task can cost seven to ten times a light one, and the same person can run both within an hour. That is precisely why per-seat intuition fails for Cowork: your cost is governed by the mix of light, medium, and heavy work across your people, not by how many people you have. Forecasting that mix by user type is the heart of the enterprise budgeting approach.

How model choice swings the per-task cost

Model use is one of the four cost inputs, and it is the one you can most directly steer. The same task costs more on a frontier model than on a lighter one. Cowork runs on Anthropic Opus 4.8 and Sonnet 4.6 at general availability; the Frontier program adds GPT 5.5; and Microsoft's fine-tuned Cowork 1, built for lower-cost everyday work, is arriving soon. Microsoft's published per-task estimates assume Opus 4.8 — the most capable and generally most expensive option — so they sit at the higher end of what is achievable.

For a buyer, this means the per-task cost is partly a policy choice. A task that does not need frontier reasoning should not run on the frontier model. Routing routine work to Sonnet 4.6 or, soon, Cowork 1 lowers the model-use component without changing the outcome that matters. The discipline to build is a routing guideline: define which categories of work justify the expensive model and which default to the cheap one, before users develop habits that quietly inflate every task. Where your tenant exposes a model picker, treat it as a budget control.

Estimating per-task volume for your people

To move from per-task cost to a monthly number, you need task volume. Start with the share of your Copilot-licensed population that will actively use Cowork in a given month — rarely everyone, especially early. For each active user, estimate a daily task count and a rough split across the three bands. A corporate knowledge worker might run mostly light tasks with the occasional medium one; a technical worker might run several heavy, multi-tool tasks a day. Apply the per-task cost to each band, multiply by working days, and sum.

That four-line model gets you within range, and you refine it as real usage data arrives during the grace period most Frontier tenants enjoy until 1 July 2026. Whether the resulting number beats a flat per-seat alternative is the question we work through in Cowork versus flat Copilot pricing, and the same per-task figures feed directly into the cost calculator.

Turning per-task cost into a controlled budget

The per-task cost is controllable in ways the headline rate is not. Prompt quality affects runtime, scope affects context retrieval, model choice affects model use, and the number of tools a workflow touches affects tool calls. Each is a lever that moves the per-task number, and together they often matter more than the $0.01 rate you cannot change. That is the buyer-side opportunity: Cowork's variability is also its controllability.

Getting there takes a governance plan and, usually, a negotiation on the seat floor and commitment terms underneath the meter. That is the work we do as independent Copilot and AI licensing advisors — modelling the per-task cost against your real workloads, then making sure the agreement around it protects the buyer. If you want a defensible per-task number for your estate, request a free Cowork cost review and we will build it with you.

Where per-task costs hide surprises

Two things inflate a task's cost in ways that do not show up in the prompt. The first is runtime spent on retries: because runtime is metered, a task that stalls, loops, or has to work hard to satisfy a vague instruction accrues credits the whole time. A poorly scoped heavy task is not just slow — it is expensive in a way a clean one is not. The second is context sprawl. Once Cowork can reach across SharePoint, Teams, email, and line-of-business systems, a broadly worded task may retrieve far more than it needs, and you pay for every source it touches. Tight scoping is a direct cost control.

There is also an attribution problem that bites at month end. Because the same user runs light and heavy tasks interchangeably, a single power user can account for a disproportionate share of a group's bill without anyone noticing until the invoice arrives. This is why per-task visibility and reporting — covered in the cost management guide — matter as much as the per-task rate itself. You cannot manage a cost you cannot attribute, and Cowork's variability makes attribution essential rather than optional.