Copilot Credits Explained: How Microsoft Meters Cowork

A Copilot Credit is the unit Microsoft uses to bill Copilot Cowork, and at pay-as-you-go pricing one credit costs exactly $0.01. That sounds simple until you ask the only question that matters for a budget: how many credits does a task consume? This page answers that, because understanding Copilot Credits is the foundation of every Microsoft Copilot Cowork cost estimate you will build.

Credits exist because Cowork does variable work. A flat per-seat fee cannot fairly price a task that reads four thousand files and runs for minutes alongside one that summarizes a single email. So Microsoft abstracts the underlying compute into a single currency — the credit — and charges you for how much of it each task burns. The credit is the meter reading; the four inputs below are the meter itself. For the wider picture of how this meter fits the total bill, start with the Microsoft Copilot Cowork cost guide, then come back here for the mechanics.

What one Copilot Credit costs

Under pay-as-you-go (PayGo), Microsoft prices a Copilot Credit at $0.01. That is the headline rate, and it is also the least negotiable number in the entire model — the money moves elsewhere, in how many credits you consume and how many seats you license underneath. The alternative payment route, P3, trades a committed usage volume for a discount on that rate, which we cover in the enterprise budgeting guide.

Because a credit is one cent, the arithmetic is easy once you know the credit count: multiply credits by $0.01 and you have the dollar cost of a task. The hard part is the credit count, and that is set per task by four inputs.

The four inputs that drive Copilot Credits

Microsoft computes a task's credit cost from four components. Each one corresponds to real work the agent performs on your behalf:

• Model use. Which model ran the task, and how hard it worked. Cowork runs on Anthropic Opus 4.8 and Sonnet 4.6 at general availability, with GPT 5.5 in the Frontier program and the lower-cost Cowork 1 arriving soon. A frontier model applying deep reasoning consumes far more than a lightweight model on a simple request — which makes model choice a genuine cost lever, not a technicality.
• Context retrieval. How much of your organizational data the task pulled in through the Work IQ context engine — documents, messages, records. A task grounded in one file is cheap; one that aggregates across SharePoint, Teams, email, and business systems is not.
• Tool calls. How many tools, plugins, or connectors the task invoked. Each call to a plugin, the Edge browser, or a Dynamics or Fabric connector is metered work that adds credits.
• Runtime. How long the task ran on Azure AI infrastructure. Long-running, multi-step jobs accrue more runtime, and runtime is compute you pay for — including time spent when a task retries or wanders because the prompt was vague.

The consequence that surprises finance teams is that two prompts which look identical can cost different amounts, because the four inputs vary with what the task actually does. This is why a credit is, in Microsoft's own framing, an abstraction over several moving parts. We translate those parts into the practical light, medium, and heavy bands in the cost-per-task guide.

How Copilot Credits convert to dollars

Put the rate and the bands together and the math becomes concrete:

A single user running ten light tasks a day across twenty working days consumes roughly 2,000-6,000 credits a month — $20 to $60. Swap several of those for heavy tasks and the same user can double or triple that figure. Multiply across thousands of users and you see why the credit count, not the per-credit price, governs the bill.

The 200-credit default every admin should revisit

Microsoft's default tenant policy caps each Copilot-licensed user at 200 credits per month — about $2 of usage — until an administrator changes it. That is a default guardrail, not a separate pool of free credits, and it is deliberately conservative because Microsoft knows uncapped agentic usage runs hot. For a buyer, 200 credits is the first setting to examine: too low and Cowork is useless to power users; removed without thought and you have handed thousands of people a metered credit card. How to scope that cap sensibly by group is the subject of the cost management and controls guide.

How Copilot Credits compare to Microsoft's other meters

Cowork's credit is not Microsoft's only consumption unit, and buyers already running other AI products will want to keep them straight. Copilot Studio meters agent activity through messages and the CCCU and ACU units across pay-as-you-go and capacity packs. Agent 365 bills per agent, a different axis entirely. Cowork adds Copilot Credits on top of the per-seat Copilot license.

The risk is paying twice for overlapping agentic work measured in different currencies. An enterprise can run Copilot Studio agents billed in CCCU, Agent 365 agents billed per agent, and Cowork tasks billed in credits, all touching the same data and the same users. Reconciling those meters against each other — and against the seat licenses beneath them — is exactly the kind of work that justifies retaining independent Microsoft licensing experts through the 2026 transition. Microsoft has little incentive to make the comparison easy.

Why the credit abstraction matters for buyers

A credit hides compute behind a clean unit, and that abstraction cuts both ways. It makes the bill legible — every task has a credit cost you can see — but it also makes the underlying margin opaque. You are not told the cost of the compute, only the credits assigned to it, and Microsoft sets the conversion. That is a reason to instrument your own usage from day one, capture real credit consumption on real workloads, and treat any vendor cost claim as a hypothesis to test rather than a number to bank.

The practical takeaway is that credits are controllable. Model choice, prompt quality, the scope of context, and the number of tools a workflow touches all move the credit count, and all are at least partly within your control. Before you turn Cowork on across a tenant, it is worth modelling that count against your seat floor and your agreement — the work we do as Copilot and AI licensing advisors, and the fastest way to turn a variable meter into a budget you can defend. When you are ready to put numbers against your own estate, the cost calculator runs this credit math in your browser, or talk to our team directly.

A worked Copilot Credits example

Take a real-world medium task: preparing a competitive briefing that pulls from a dozen internal documents, two connected market-data plugins, and a structured reasoning pass to produce a summary and a slide outline. The model-use input reflects a capable model applying structured reasoning; context retrieval reflects a dozen sources; tool calls reflect the two plugins invoked several times; runtime reflects a job that runs for a few minutes. Summed, the task lands in the medium band — call it 450 credits, or $4.50.

Now run the same briefing differently. Narrow the source set to the five documents that actually matter, route it to a lighter model because the reasoning is not frontier-grade, and write a tighter prompt so the agent does not retry. The model-use, context, and runtime inputs all drop, and the same useful output now costs perhaps 220 credits, or $2.20 — better than a 50% reduction for an essentially identical result. That gap, repeated across thousands of tasks a month, is the difference between a controlled Cowork budget and a runaway one. It is also why we treat prompt design and model routing as cost controls, not just quality settings, when we help clients operationalize Cowork.