Copilot Studio Pricing: Microsoft Copilot Studio Pricing Model, Plans, Credits and Agent Costs

Copilot Studio prices consumption, not seats. There is no published per-user monthly fee for the product itself, and the maker licence that lets somebody build an agent costs nothing. What you buy is Copilot Credits, a single currency that measures how much work your agents do, pooled at the tenant level and drawn down by every agent anyone in your organization publishes.

Credits replaced the old unit, "messages", on September 1, 2025. Microsoft kept the pack quantity and the pay-as-you-go rate identical through that change, so a quote written against messages is still numerically right about the pack, but wrong about how usage is counted: credits are charged per feature invoked, and one user turn can fire several features at once.

There is a sequencing detail that catches people out on day one. The free Copilot Studio user licence is not something an admin can simply hand out. Microsoft documentation is explicit that you need a tenant prepaid Copilot Credit pack subscription before you can get that user licence. The free part is the seat, not the entry ticket.

Three things then decide your monthly number, in descending order of impact: how many people talk to the agent, what each interaction makes the agent do, and whether those people hold Microsoft 365 Copilot licences.

These are the published rates as of August 2026. They apply to the language models Copilot Studio provides and exclude bring-your-own-model configurations such as Azure Foundry models, which bill separately.

At the prepaid rate of roughly $0.008 a credit, an agent action costs about four cents, a tenant graph grounded answer about eight, and a minute of premium generative voice about sixty.

Reasoning models bill twice. This is the line most cost models miss. When an agent uses a reasoning-capable model, Microsoft charges the feature rate for the operation plus the premium text and generative AI tools rate of 10 credits per 1,000 tokens for the reasoning itself. A generative answer from a reasoning model is therefore 2 credits plus 10 credits for every 1,000 tokens it thinks through, not 2 credits. Deep reasoning is the single easiest way to turn a cheap agent into an expensive one.

One more exclusion worth flagging: Computer-Using Agents bill at the agent action rate of 5 credits, but they are not covered by the Microsoft 365 Copilot inclusion. Everything else in that column becomes free for a licensed employee. Computer use does not.

This is the part that makes most published Copilot Studio pricing advice out of date, and it is worth reading before you trust any estimate you find online.

Everything you build in Copilot Studio runs on a runtime that Microsoft calls a harness, sitting between your design and the model. There are now three of them, and Microsoft documentation states plainly that they use different billing methods.

The practical consequence is a budgeting trap. If your team picks the GitHub Copilot harness because it is the capable one, and it is, the proof-of-concept has a running cost from the first prompt. Credits cover the model tokens, the tools including knowledge sources and MCP servers, and the harness itself. On the standard harness you can prototype for nothing and only start paying at publish. Two teams doing identical discovery work on the same idea can get very different first invoices depending on which runtime they clicked.

It also means you should ask any vendor, consultant, or blog quoting you a Copilot Studio number which harness they costed. If they cannot answer, the number predates the split.

The floor is one prepaid capacity pack: 25,000 Copilot Credits a month, widely listed at $200. Above that, the honest answer is that the agent design sets the price, and the spread is enormous. Microsoft publishes three worked examples in its own billing documentation, and they are the fastest way to see the range.

Read those three rows together and the rule falls out: conversation volume is what costs money, back-office automation is close to free. The support agent burns more than eight capacity packs a month. The order processing agent burns about a quarter of a percent of one. Same product, same rate card, a 350x difference in cost, driven entirely by how many humans are talking to it and how deeply each turn is grounded.

Notice what the sales performance example quietly shows about licensing, too. It has 150 users but bills for 100, because the 50 Microsoft 365 Copilot licensed users are zero-rated. That single fact is often worth more than any negotiation on the pack price.

Partly, and the boundary is precise enough to plan around. A Microsoft 365 Copilot licence at $30 per user per month lets that user extend Microsoft 365 Copilot with agents, and it zero-rates a specific list of things when the agent runs under that user authenticated identity.

What becomes no charge for a Microsoft 365 Copilot licensed user: classic answers, generative answers, agent actions, tenant graph grounding, and agent flow actions triggered specifically by "When an agent calls the flow". Microsoft describes this as business-to-employee usage included in the user subscription licence, subject to fair usage limits it reserves the right to revise.

What stays chargeable: anything customer-facing, anything invoked by an unlicensed user, Computer-Using Agents, agent flows fired by any other trigger, and bring-your-own-model configurations. Agents built in Agent Builder inside Microsoft 365 avoid generative answer charges only when the response does not use tenant graph grounding.

So the same agent genuinely has two price tags. Point it at your own staff who already hold Copilot licences and it is close to free to run. Point it at your customers and every single turn meters. If you are weighing the seat cost itself, the full breakdown sits in our guide to Microsoft 365 Copilot pricing, and the credit mechanics are covered in more depth in Copilot Credits pricing.

Microsoft does not throttle you. It switches the agent off. This is the single most important operational fact in Copilot Studio pricing and it deserves more attention than it usually gets.

Enforcement triggers when a tenant reaches 125% of prepaid capacity. Microsoft allows that 25% as a grace band so a business process does not break the moment a pack empties. Past it, custom agents are disabled. A conversation already in flight finishes normally, then the agent stops responding and users get one of two messages: "There is a billing issue." or "This agent is currently unavailable. It has reached its usage limit." An email goes to the tenant administrator and a notice appears in the Power Platform admin center.

Agent flows are stricter. Flow enforcement kicks in when prepaid capacity is fully consumed, at 100%, not 125%. New flow runs are rejected, runs already in progress finish, and the parent agent keeps working normally for classic answers, generative answers, and agent actions. Test runs in the designer are exempt. Worth separating in your head, because it means the visible failure and the silent failure happen at different thresholds.

Three things blunt this risk:

• Link the environment to pay-as-you-go. Environments on the Azure meter are exempt from enforcement entirely, because overage simply bills. If an agent faces customers, this is the setting that stops it going dark mid-quarter.
• Cap individual agents. You can set monthly consumption limits per agent in the Power Platform admin center under Licensing, Copilot Studio, Manage Agents, so one runaway agent cannot eat the tenant pool.
• Allocate capacity per environment. Reserved environment capacity is honored, so a production environment with its own allocation keeps running even when the shared tenant pool has already tripped enforcement.

And remember the reverse problem: unused credits expire monthly. Buying enough capacity to survive your busiest month means paying for headroom you burn in eleven other months. Prepaid for the baseline plus pay-as-you-go for the peaks is usually cheaper than prepaid for the peak.

Yes. Microsoft publishes a free Copilot Studio agent usage estimator at microsoft.github.io/copilot-studio-estimator, and it is the right starting point because it models the variables that actually drive the bill rather than asking for a headcount. You select agent type, expected traffic, orchestration style, knowledge sources, and tools, and it returns a credit forecast you can convert straight to dollars at $0.008 prepaid or $0.01 pay-as-you-go.

Three inputs deserve care, because a small error in any of them moves the answer by an order of magnitude:

• Turns per conversation, not conversations. Microsoft support example costs 8 credits per conversation because it counts six answers, not one. Estimating per conversation instead of per turn is the most common way people undercount by 5x.
• Tenant graph grounding on or off. It is a per-agent toggle and it is 10 credits, five times a generative answer. Leaving it on for an agent that does not need tenant-wide retrieval is the most expensive default in the product.
• Whether a reasoning model is in play. If it is, add the premium token meter on top of the feature rate. The estimator handles orchestration choices, but you should sanity-check reasoning token volume yourself.

Then apply the licence filter. Split your forecast volume into employee-facing turns from Microsoft 365 Copilot licensed users, which are zero-rated, and everything else, which is not. For a lot of internal-only agents that step removes most of the estimated cost. For a customer support agent it removes none of it, and that is exactly the case where buying a finished agent on a flat subscription tends to beat metered building.

Both are legitimate, and most organizations end up running both. The decision is usually settled by three questions, not by preference.

Who talks to it? If the answer is your own staff and they already hold Microsoft 365 Copilot licences, building in Copilot Studio is very hard to beat on price, because the usage is zero-rated. If the answer is customers, prospects, or unlicensed frontline workers, every turn meters and the economics change completely.

Is the process yours or is it generic? An approval chain shaped by your own policies, running on your own Dataverse tables, is exactly what Copilot Studio is for. A support triage agent, a lead qualifier, or a document extraction step is not company-specific, and building one from scratch means paying build credits, engineering time, and ongoing maintenance for something that already exists finished.

Can you tolerate a hard stop? Prepaid capacity plus a 125% cutoff is a real operational risk for anything customer-facing. A flat monthly subscription has no pool to exhaust.

Where a ready-made agent wins is the middle case: a well-understood job, meaningful volume, and a team that would rather spend its Copilot Credits on the workflows only it can build. Agents in our catalog start at $29 a month, connect to your stack through scoped, revocable API permissions, and carry no credit pool, no monthly expiry, and no usage cliff. Every listing shows its capabilities, integrations, permissions, and limits up front, so you can see exactly what it does before you deploy it. Browse the agent catalog, or compare the major vendor catalogs side by side in our guide to AI agent marketplaces.

A reasonable split for most teams: build the workflows that encode your own rules in Copilot Studio, keep the credit pool sized for those, and buy the commodity jobs.