The short version: A Copilot Studio tenant capacity pack buys 25,000 messages a month for USD 200, as at 22 September 2026. Pay-as-you-go runs USD 0.01 a message with no commitment. A message isn't one fixed thing: a scripted answer costs 1, a generative answer costs 2, and grounding it in your own tenant data adds 10 more on top of that.
25,000 messages, USD 200 a pack
| Setting | Value | Source |
|---|---|---|
| Tenant capacity pack size | 25,000 messages a month | Microsoft Copilot Studio pricing |
| Capacity pack price | USD 200 a month per pack | Microsoft Copilot Studio pricing |
| Pay-as-you-go rate | USD 0.01 a message, no commitment | Power Platform pay-as-you-go meters |
| Unused pack messages | Don't roll over to next month | Microsoft Copilot Studio capacity packs |
Copilot Studio is licensed by tenant, not by seat: buy one pack and every agent, in every environment, draws from the same pooled 25,000 messages a month, as at 22 September 2026. A prepaid pack works out to USD 0.008 a message, about 20% cheaper than pay-as-you-go, but the trade is that unused messages don't carry over, so a pack only pays for itself once usage has settled into something predictable. For the full arithmetic on sizing a pack against a real workload, I've worked a small-firm example in the Copilot Studio pricing breakdown.
What counts as a message
Not every reply costs the same, which is where a pack that looked generous on paper runs out faster than expected. A scripted, classic answer, one following a path a maker built in the authoring canvas, costs 1 credit. A generative answer, composed by AI from a knowledge source, costs 2. Ground that answer in your own tenant data through Microsoft Graph, SharePoint, Teams or Outlook, and it adds 10 more on top of the generative cost. An agent action, such as calling a connector to look something up or complete a task, costs 5. These stack inside a single exchange, so a question that gets a generative answer grounded in tenant data costs 12, six times a plain scripted reply. I've broken down where a real deployment burns through a monthly allowance fastest in the credit meter breakdown.
One exemption is worth knowing before you size a pack: a Microsoft 365 Copilot seat draws on the seat, not the pack, for an agent's classic answers, generative answers, agent actions and tenant graph grounding used inside Copilot Chat, Teams or SharePoint. Autonomous, event-triggered runs and anything answering an external customer don't get that exemption, and still bill against the pack or the pay-as-you-go meter.
Sizing a pack for a firm under 20 staff
Most firms under 20 staff building their first internal agent never come close to 25,000 messages a month. A team of 15 people asking an agent a handful of questions a day, weighted toward simple scripted and generative answers rather than heavy tenant-wide grounding, typically lands in the low hundreds of messages a month, comfortably inside pay-as-you-go territory at a few dollars. The pack starts making sense once an agent is answering customers directly, running autonomously off a trigger, or grounding most of its answers in tenant data by default rather than a scoped knowledge source, since each of those pushes the per-message cost up several times over. Size the decision on your own traffic rather than the sticker price of the pack: I model both options with real numbers as part of a Workflow Review, before anyone commits to either.
Common questions
How many messages does a Copilot Studio capacity pack include?
25,000 a month, for USD 200. It's licensed by tenant, not by seat, so every agent in every environment in your tenant draws from the same pooled 25,000, however many people are using them.
What counts as one message in Copilot Studio?
It depends what the agent did. A scripted, classic answer costs 1 credit. A generative AI answer costs 2. Grounding that answer in your own tenant data through Microsoft Graph adds 10 more on top, and an agent action such as a connector call costs 5. They stack inside a single reply, so one complex question can cost ten times a simple one.
Does a Microsoft 365 Copilot licence use up the message pack?
Not for the core interactions. When someone with an M365 Copilot seat uses an agent inside Copilot Chat, Teams or SharePoint, their classic answers, generative answers, agent actions and tenant graph grounding are zero-rated and don't draw from the capacity pack. Autonomous, event-triggered runs and anything reaching an external customer don't get that exemption.
Not sure whether you need a pack at all?
Free 30-minute call. Tell me what you're planning to build and I'll tell you whether pay-as-you-go or a pack is the right call.