Copilot Studio pricing, cost per Copilot Credit, and what Microsoft 365 Copilot agents actually bill you.

What a Copilot Credit is worth, in plain dollars

Microsoft sells Copilot Studio as tenant-wide packs of 25,000 Copilot Credits priced at $200 per pack per month. Divide one by the other and a single Copilot Credit is worth eight tenths of a cent, $0.008. That one conversion is the key to the entire rate card, because everything else Microsoft publishes about agent cost is denominated in credits rather than dollars, and almost nobody does the division for you.

The capacity is pooled across the whole tenant rather than assigned per user or per agent, which is genuinely convenient. It also means a single badly-behaved agent in one environment can eat the budget belonging to every other agent, which is why Microsoft provides per-environment capacity allocation. Pay-as-you-go is the same product billed differently: you link an environment to an Azure subscription through a billing policy in the Power Platform admin center and pay for the credits you actually consumed at the end of the month. Microsoft's own FAQ is clear that there are no feature or capability differences between the pack and the meter, only the timing of the payment.

The full Copilot Credits rate card, converted to dollars

Here is the published rate table with the arithmetic done. A classic answer, meaning a response an author wrote by hand, costs 1 credit, so $0.008. A generative answer costs 2 credits, so $0.016. An agent action, which covers triggers, deep reasoning, topic transitions and computer-using agents, costs 5 credits, so $0.04. Tenant graph grounding for messages costs 10 credits, so $0.08. Agent flow actions are billed at 13 credits per hundred actions, which is about a tenth of a cent each and by far the cheapest line on the card. Content processing tools cost 8 credits a page, so $0.064, meaning a thousand-page document batch runs about $64.

The prompt tools are where the spread gets wide. Basic text and generative AI tools cost 1 credit per ten responses, standard cost 15 credits per ten, and premium cost 100 credits per ten. In dollars that is $0.008, $0.12 and $0.80 for the same ten responses, a hundredfold range depending purely on which underlying model the prompt uses. Voice is metered by the minute across three tiers: classic voice at 10 credits a minute, generative voice at 35, and premium generative voice at 75. That last one is $0.60 a minute, which is $36 an hour, and it means one $200 credit pack buys about five and a half hours of premium voice a month before anything else on the agent has run at all.

For context against the other enterprise agent platform buyers usually shortlist alongside this one: a Salesforce Flex Credit is worth $0.005 rather than $0.008, but a standard Agentforce action costs 20 credits against 5 for a Copilot Studio agent action. In dollars that is $0.10 against $0.04, so Copilot Studio is the cheaper of the two per unit of agent work. The two platforms differ far more sharply on what happens when you overspend, which we set out in Agentforce vs Copilot Studio.

Copilot Credits to tokens: what the prompt meter costs per million

Microsoft's billing rates table publishes a per-token rate for prompt tools that almost nobody converts, and it is the clearest way to compare Copilot Studio against a raw model API. The rates are 0.1 Copilot Credit per 1,000 tokens for basic text and generative AI tools, 1.5 credits per 1,000 tokens for standard, and 10 credits per 1,000 tokens for premium. Those are the same three tiers quoted elsewhere as 1, 15 and 100 credits per ten responses; this is the same meter expressed per token instead of per response.

Multiply by $0.008 a credit and the picture gets sharp. Basic prompt tools cost $0.0008 per 1,000 tokens, which is $0.80 per million tokens. Standard costs $0.012 per 1,000, which is $12 per million. Premium costs $0.08 per 1,000, which is $80 per million tokens. That hundredfold spread between basic and premium is decided entirely by which model tier a prompt tool uses, and it is set inside the agent rather than at purchase time, so it is not something your procurement conversation will surface.

The $80 per million figure is the one to carry around, because premium is what applies when an agent uses a reasoning model, and reasoning is exactly what teams switch on late in a pilot. Compared against commodity model pricing, $80 per million tokens is expensive; compared against the engineering you would otherwise do to ground, secure and govern the same agent inside a Microsoft tenant, it may well be worth it. The point is that the comparison is now possible, which it is not while the rate sits in credits.

How to buy Copilot Credits: capacity packs, pay-as-you-go and commit units

Microsoft sells the same credits three different ways, and which one you pick changes the price rather than the product. The first is the prepaid Copilot Credit capacity pack: 25,000 credits for $200 a month, bought as a tenant subscription, billed whether you use them or not. The second is pay-as-you-go, where you link an environment to an Azure subscription through a billing policy in the Power Platform admin center and pay for what you actually consumed at the end of the month. The third, and the one almost nobody writes about, is a Copilot Credit prepurchase pool bought in the Azure portal as a one-year commitment, which Microsoft prices at a saving of up to 20 percent against the standard rate.

That third route is the one worth knowing if your usage is both large and predictable, because it is the only mechanism Microsoft offers that reduces the unit price at all. Everything else on this page is priced at $0.008 a credit; a commit unit pool is the single lever that moves that number. The tradeoff is the usual one: you are committing for a year against a consumption forecast, and Copilot Studio consumption is famously hard to forecast, particularly if reasoning models are anywhere in your roadmap. Buy the commitment after you have measured a real month, not before.

Choosing between the first two is simpler than it looks, because Microsoft states plainly that there are no feature or capability differences between the pack and the meter, only the timing of payment. So the decision is purely about predictability. Steady, known volume favors packs, because you get a fixed bill and you can allocate capacity per environment. Spiky or unknown volume favors pay-as-you-go, because you are not paying for headroom you did not use, and because pay-as-you-go does not hit the enforcement wall that disables custom agents on prepaid capacity, and Microsoft's own Copilot Credits Guide says the meter carries no interruptions due to overages. Many tenants end up running both: packs for the baseline, the meter for the overflow.

The 3,750 credit rule, which decides seats against the meter

This is the piece of arithmetic the roundups never do, and it is the single most useful number in this whole cluster. Microsoft 365 Copilot costs $30 per user per month paid yearly, and a Microsoft 365 Copilot license zero-rates employee-facing agent usage: classic answers, generative answers and tenant graph grounding all show as "no charge" in the billing rates table when the user is licensed and the agent runs under that user's authenticated identity. So the question is not really "packs or pay-as-you-go", it is "seats or meter", and there is a clean breakeven.

At $0.008 a credit, $30 buys exactly 3,750 Copilot Credits. That is the line. If an employee generates more than 3,750 credits of agent activity in a month, the $30 Microsoft 365 Copilot seat is cheaper than metering them; below it, the meter wins. Translated into things people actually do, 3,750 credits is 1,875 generative answers, or 750 agent actions, or about 312 tenant graph grounded questions. Spread across roughly 21 working days, that last figure is about 15 grounded questions a day. So the rule of thumb is this: an employee who asks a tenant-grounded agent more than about fifteen questions a working day has already justified a Microsoft 365 Copilot seat on agent usage alone, before you count anything the license does inside Word, Excel or Teams.

Two caveats keep this honest. The zero-rating is explicitly limited by fair usage limits that Microsoft reserves the right to change, and it does not apply to everything. Computer-using agents are excluded by name. Agent flows are only zero-rated when triggered by "When an agent calls the flow", and any other trigger bills at the standard rate. And none of it applies to customer-facing agents, because the person on the other end does not hold a license.

What Microsoft's own worked examples actually cost per month

Microsoft publishes three billing examples in its documentation, all quoted per day. Converting them to monthly bills is instructive, and it is the fastest way to understand why the sticker price of $200 is misleading.

The first is a customer support agent on a website answering questions from return policies and product manuals. Microsoft models an average run as four classic answers and two generative answers, which is 8 credits, across 900 customers a day, giving 7,200 credits a day. At 25,000 credits a pack, Microsoft's own example agent exhausts a $200 credit pack in under three and a half days. Over a thirty-day month it consumes 216,000 credits, which is nine packs, or roughly $1,728 to $1,800 a month for one website support agent. That is the number worth carrying away from this page.

The second is a sales performance agent grounded in the tenant graph, serving 50 Microsoft 365 Copilot licensed users and 100 unlicensed ones. Only the unlicensed users bill, at 48 credits each per run, giving 4,800 credits a day or about $1,152 a month. The 50 licensed users cost nothing in credits, though they do cost $1,500 a month in seats, which is the seats-against-meter tradeoff in live form. The third is an internal order processing agent that fires four actions per order at 5 credits each, so 20 credits, or 16 cents per order processed. That one is genuinely cheap, and it shows the pattern: narrow autonomous agents doing a fixed number of actions are inexpensive, while open-ended conversational agents at consumer volume are not.

The free license that is not free to reach

"Is Copilot Studio free" is one of the most common searches in this cluster, and the honest answer has three parts. There is a trial license anyone can sign up for individually, but the documentation states plainly that you can build and test agents with it and cannot publish them, which rules out any production use. There is a Copilot Studio user license that is genuinely free of charge and assigned by an admin. And there is the catch: the same page notes you need a Copilot Studio tenant prepaid Copilot Credit pack subscription before you can get that user license.

So the free license is real, but it is downstream of a $200 a month purchase. This is the same structural pattern we found when we read Salesforce's Agentforce rate card, where a $5 per user Agentforce User License still requires Flex Credits to do anything, and the same pattern behind GoHighLevel AI Employee pricing, where the AI is an add-on priced per sub-account on top of a platform plan. It is worth recognizing as a category: in agent products, the cheap or free license is almost never the thing that makes the agent run.

Reasoning models bill twice, and that is where budgets break

This detail sits in one paragraph of Microsoft's billing documentation and it deserves far more attention than it gets. When an agent uses a reasoning-capable language model, Copilot Studio bills on two meters simultaneously. The feature rate covers the core action, a generative answer or an agent action, and it always applies. On top of that, the premium text and generative AI tools rate applies to the reasoning model's token usage at 10 credits per thousand tokens. Microsoft states the formula outright: total cost equals the feature rate plus the premium tools rate.

Run the numbers and the effect is dramatic. A generative answer on a standard model costs 2 credits, $0.016. The same answer from a reasoning model that consumes three thousand tokens costs 2 credits for the feature plus 30 credits for the tokens, so 32 credits, about $0.256. That is sixteen times the price for what the user experiences as the same answer, arriving slightly slower. Anyone piloting an agent on a standard model and then switching on deep reasoning before rollout should expect the bill to change by an order of magnitude, and should re-run their estimate rather than trusting the pilot. Microsoft does publish an agent usage estimator, which is the right tool for this, but the estimator is only as good as the assumption you feed it about how often reasoning fires.

What happens when you run out, which is harsher than most meters

Most consumption products degrade politely when you exceed capacity. Copilot Studio does not. The documentation sets out an enforcement policy that applies once an environment uses all its allocated credits, at which point custom agents are disabled. Microsoft allows some overage first, and it does put a number on it: the Overage enforcement section of its billing rates and management article, updated on August 3, 2026, states that enforcement is triggered when a tenant reaches 125 percent of its prepaid capacity, at which point custom agents are disabled. The threshold is applied to tenant-level capacity rather than to each environment separately, and Microsoft works through a four-environment example to show the difference. An ongoing conversation is allowed to finish, and every subsequent attempt to invoke the agent is rejected until capacity is increased or the month resets.

The part worth reading twice is what the end user sees. Microsoft documents two responses: "There is a billing issue." and "This agent is currently unavailable. It has reached its usage limit." If the agent in question is the support agent on your public website, those are the sentences your customers read. An email goes to the tenant admin and a notice appears in the Power Platform admin center, but the customer-facing failure is not silent. Agent flows behave slightly differently, blocking new runs rather than disabling the agent outright. Combined with the fact that unused credits do not carry over, the practical guidance is to allocate capacity per environment so that a runaway agent cannot take down the ones that matter, and to set the alerting up before launch rather than after the first incident.

Where Copilot Studio is clearly the right buy

If your organization already runs Microsoft 365 and your people already have Copilot licenses, this is not a close call. The zero-rating alone changes the economics: employee-facing agents grounded in your own tenant data cost nothing extra for licensed users, which is a genuinely strong offer and one no outside tool can match. Add the admin tooling, the Entra-based access control, the per-environment capacity allocation and the fact that the agent sits where the work already happens, and Copilot Studio is the obvious answer for internal agents in a Microsoft shop. If you need voice, it is also the answer, because we do not do voice at all.

The case for us is narrower and worth stating precisely. You are a small business or a founder-led team. You do not run a Microsoft tenant, or you do but you have no appetite for Azure billing policies, capacity allocation and an admin center. You want a defined outcome rather than a platform to build on: research done, a lead list built, outreach drafted, an inbox triaged, a report produced. And you want the bill to be the same in March as it was in February. That is what a flat-rate agent is for, and it is why our price is one number.

How to work out your own number in fifteen minutes

Start with seats, because that decision dominates everything else. Count the people who will actually use agents. If they will be heavy internal users, meaning more than roughly fifteen tenant-grounded questions a working day each, price Microsoft 365 Copilot seats at $30 per user per month and treat their agent usage as included, checking first whether the cheaper business add-on covers them on our Copilot license cost page. If usage will be light or occasional, leave them unlicensed and plan to meter them.

Then estimate consumption, and only for the traffic that actually bills: customer-facing agents, unlicensed internal users, computer-using agents and non-standard agent flow triggers. Take your monthly interaction volume, multiply by a realistic credits-per-interaction figure using Microsoft's examples as a sanity check (8 credits for a simple support run, 48 for a tenant-grounded one, 20 for a fixed four-action automation), and multiply by $0.008. If Salesforce is the other half of your comparison, the equivalent sum is already built into our Agentforce pricing calculator. Round up to whole packs at $200 each, or plan to use pay-as-you-go if the volume is genuinely unpredictable. The Copilot Studio credits calculator does that rounding for you, and the purchasing routes themselves are set out in how to buy Copilot Credits. Finally, add a reasoning multiplier if you intend to use deep reasoning at all, and add capacity headroom, because the grace period is 25 percent and enforcement at 125 percent disables the agent rather than simply invoicing you for the excess. Compare that total against $149 flat and the answer will be obvious in either direction, which is the point of doing the arithmetic rather than reading a listicle.