← All insights

Architecture guideLens: United States5 min read

Copilot Studio pricing explained: Copilot Credits, licenses and what an agent really costs to run

How Copilot Studio is priced in 2026: credit packs, pay-as-you-go and pre-purchase plans, what Microsoft 365 Copilot includes, the rate card, reasoning models and the limits that block agents.

Two ancient bronze coins on a white background, in La Madre duotone, beside the words Count the credits
Photo: Amisos / BNF (Wikimedia Commons, CC0)

“How much does Copilot Studio cost?” has a short answer and a useful one. The short answer: you pay for Copilot Credits, and a pack of 25,000 credits costs US$200 a month. The useful answer is that two agents with the same number of users can differ in cost by a factor of ten, because Copilot Studio meters what an agent does, not who uses it.

This guide covers how to buy credits, what Microsoft 365 Copilot already includes, the rate card, the special case of reasoning models and the new harness, and the enforcement rules that can take an agent offline. Figures are US list prices from Microsoft’s documentation as read on November 20, 2026; prices in other currencies and under enterprise agreements differ, and Microsoft updates the rate card regularly.

Three ways to buy Copilot Credits

Capacity packs. The Copilot Studio license is a tenant-wide subscription sold in packs of 25,000 credits per month at US$200 per pack, billed annually. Capacity is pooled across the tenant and can be allocated to environments. Unused credits do not roll over to the next month.

Pay-as-you-go. You link a Power Platform environment to an Azure subscription through a billing policy, and pay at the end of each month for the credits actually consumed, at US$0.01 per credit in Microsoft’s licensing guide. No commitment, and no blocking when usage spikes.

Pre-purchase plans. The Copilot Credit Pre-Purchase Plan is a one-year upfront commitment in Copilot Credit Commit Units, where one unit is worth US$1, or 100 credits, with tiered discounts that Microsoft’s pricing page puts at up to 20%. The Microsoft Agent Pre-Purchase Plan works the same way but its units can be spent on Copilot Studio or Microsoft Foundry. Unused units expire at the end of the year. Microsoft also lists an “AI Flex Plan,” across Copilot Studio and other AI services, as coming soon.

People who build agents need either a Copilot Studio user license, which costs nothing once the tenant has capacity, or, under pay-as-you-go and pre-purchase, membership in the security group assigned as Copilot Studio authors. A trial license lets a maker build and test, but not publish.

What Microsoft 365 Copilot already includes

If the people using an agent have Microsoft 365 Copilot licenses, many employee-facing interactions cost nothing extra: classic and generative answers, agent actions, tenant graph grounding and AI tools, when the agent runs in Copilot Chat, Teams or SharePoint with the licensed user’s identity. Agent flows are included only when the agent calls them through the “When an agent calls the flow” trigger.

Three common cases are not included: agents that run autonomously, for example on an event trigger, even if everyone in the company is licensed; agents for customers or other external users; and computer-using agents. Fair-use limits apply to the included usage.

The rate card

On the Standard harness, each thing an agent does has a price in credits:

Copilot Studio billing rates, Standard harness
FeatureCopilot CreditsWith a Microsoft 365 Copilot user
Classic answer (authored response)1No charge
Generative answer2No charge
Agent action (trigger, reasoning step, topic transition, computer use)5No charge, except computer use
Tenant graph grounding10No charge
Agent flow actions13 per 100 actionsNo charge only via the agent trigger
AI tools, basic / standard / premium0.1 / 1.5 / 10 per 1,000 tokensNo charge
Content processing8 per pageNo charge
Voice, per minute10, 35 or 75 by tierNot listed
Source: Microsoft Learn, Billing rates and management, last updated October 30, 2026. Check the current page before estimating.

Microsoft’s own examples show how fast design changes the bill. A website support agent averaging four classic answers and two generative answers per conversation, with 900 customers a day, uses 7,200 credits a day, about US$72 a day at the pay-as-you-go rate. An internal agent grounded in the tenant graph, answering 100 unlicensed users with four generative and four grounded answers each, uses 4,800 credits a day, most of it from grounding.

Reasoning models and the new harness bill by tokens

When a reasoning model orchestrates an agent, Copilot Studio adds a second meter: 10 credits per 1,000 billed tokens, on top of the feature rate. Billed tokens include input, output and reasoning, and content written to the prompt cache; tokens read from the cache are free. In Microsoft’s worked example, one follow-up question to a support agent using a reasoning model costs 141 credits plus feature rates, and the same question in a new conversation costs 461, because less of the prompt is cached. Long instructions, large knowledge retrievals and long conversations all raise the bill.

Agents on the Microsoft Copilot harness, generally available since Ignite 2026, use usage-based billing: credits for model tokens, tools such as knowledge and MCP, and the harness itself. And they start consuming credits when you start building, including tests and evaluations. We covered what that means for project budgets.

The limits that can take an agent offline

On prepaid capacity, Microsoft allows some overage, then blocks agents when the tenant reaches 125% of what it bought. A blocked agent stops responding, even in the middle of a conversation, with a message that it “has reached its usage limit.” Users with Microsoft 365 Copilot licenses might not be blocked. Environments linked to pay-as-you-go are not blocked; the excess is billed instead.

Agent flows are enforced separately: when an environment has no credits, new flow runs fail while the agent keeps answering. And a monthly limit set on an agent does not stop the agent flows it calls. Power Automate cloud flows, by contrast, use Power Automate licensing and are not billed in credits.

How to estimate an agent before you build it

  1. List the interactions, not the users. For each scenario, count classic answers, generative answers, actions, grounded answers and flow actions per conversation.
  2. Split licensed and unlicensed users. Included usage only applies to Microsoft 365 Copilot users acting with their own identity inside Microsoft 365.
  3. Decide the model. If the agent does not need deep reasoning, a standard model avoids the token meter. Microsoft itself recommends this.
  4. Run Microsoft’s usage estimator, then test with real questions and compare the consumption shown in the Power Platform admin center with your estimate.
  5. Choose the purchase model last. Pay-as-you-go for uncertain or seasonal volumes, packs or pre-purchase once a month of real usage shows the baseline. Keep pay-as-you-go linked behind prepaid capacity for agents that must not stop.

For what Copilot Studio is and when to use it, start with our Copilot Studio guide. If you want an estimate for a specific agent, that is part of how we run Copilot Studio projects.

Have an AI use case stuck between prototype and production?

Tell us what you’re trying to ship. We’ll reply with honest next steps.

Discuss a use case