Skip to content

How Copilot Credits work — what an agent costs to run

Once you reach Stage 6 · Copilot Studio, your agents stop being free personal helpers and start metering usage. Copilot Studio measures that usage in Copilot Credits — and so do Copilot Cowork, Autopilot, and the Work IQ APIs. This page explains the model so the numbers in the Credit Estimator — and on your monthly bill — actually make sense.

Everyday AI vs. advanced AI — where the meter starts

Microsoft now splits Copilot into two ways of paying (USL vs. usage-based billing):

  • Everyday AI — Copilot Chat, Copilot in the Office apps, the built-in agents, and Agent Builder — is covered by the Microsoft 365 Copilot user license (USL): a fixed price per user.
  • Advanced AI — Cowork, Code, Autopilot, agentic SharePoint, and the newest frontier models — runs on usage-based billing in Copilot Credits, and still requires the USL for access.

Copilot Studio agents also bill in Copilot Credits, but with a big exception for licensed employees — see the Microsoft 365 Copilot license below.

Borrow, don't build — this is grounded, but Microsoft owns the truth

Every rate on this page is taken from Microsoft's official billing docs, re-verified 2026-10-28. Rates and packaging change; for any customer-facing number, go to the source:

The one idea to internalize: build size ≠ run cost

The single most common estimation mistake is treating "how big is this agent" and "how much will it cost to run" as the same question. They are two independent axes.

  • Build size — a T-shirt size (XS → XL) that measures the effort to design, stand up, and own the agent: how many actions and back-end systems it wires, whether it's autonomous, whether it uses voice. It's a one-time difficulty signal.
  • Run cost — the Copilot Credits per interaction × how many interactions per month. It's an ongoing, recurring number driven mostly by volume.

These move independently:

A tiny agent can outspend a complex one

A one-topic FAQ agent (XS to build) that answers 50,000 employee questions a month will burn far more credits than a multi-system, human-escalation workflow agent (L to build) that runs 200 times a month. Size tells you how hard it was to build. Only volume × rate tells you the bill.

The Credit Estimator reports these as two separate readouts on purpose — Build effort and Run cost — with the exact drivers behind each.

The unit: Copilot Credits (formerly "messages")

Copilot Credits are the common currency for Copilot Studio agent usage. You get a pool of credits with your licensing and it's shared across the whole tenant — every agent draws from the same pool.

The rename you'll still see in old material

On September 1, 2025, the currency changed from messages to Copilot Credits. Per Microsoft, there was no change to the quantity per prepaid pack or to the pay-as-you-go rate — only the name. If a blog or screenshot says "messages," read it as "credits."

The rate card — what each thing costs

Every credit-bearing capability on the standard harness, straight from Microsoft's billing rates (re-verified 2026-10-28 — unchanged since August):

Agent capability Rate
Classic answer (pre-authored, static reply) 1 credit
Generative answer (AI reads knowledge + responds) 2 credits
Agent action (a connector/tool step, deep reasoning, topic transition) 5 credits
Tenant graph grounding (RAG over your Microsoft 365 Graph) 10 credits
Agent flow actions 13 credits per 100 actions (~0.13 each)
Text & generative AI tools — basic 1 credit per 10 responses
Text & generative AI tools — standard 15 credits per 10 responses
Text & generative AI tools — premium (advanced reasoning) 100 credits per 10 responses
Content processing tools 8 credits per page
Voice — classic / GenAI / premium GenAI 10 / 35 / 75 credits per minute

A few things this table teaches:

  • Credits stack per interaction. A single complex prompt can hit several meters at once. Microsoft's own worked example: an agent grounded in the tenant graph uses 12 credits for one response — 10 for tenant graph grounding + 2 for the generative answer.
  • Voice is priced by the minute, not the turn, and the per-minute rate already includes the core classic/generative/action activity during the call — so you don't add those on top.
  • Reasoning models bill twice. A reasoning-capable model charges the normal feature rate plus the premium "Text & generative AI tools" meter — 10 credits per 1,000 tokens — for the extra deep-reasoning compute. Tokens the model reads back from its prompt cache aren't billed, so concise instructions and fresh conversations for new topics keep this down.
  • Grounding on your own documents is not a separate line. Answering over a SharePoint site or a set of files is part of the generative answer — it's the tenant graph grounding that adds the 10-credit meter.

Volume is the multiplier — and autonomous ≠ interactive

The rate card is only half the story; volume is what actually sets the bill, and how you count volume depends on the agent's regime:

  • Interactive agents are driven by people: cost ≈ users × interactions per user per month × credits per interaction. Reach and frequency dominate.
  • Autonomous (triggered) agents are driven by events — a file lands, a message arrives, a schedule fires. Cost ≈ events per month × credits per event. There's no per-user framing, so the Microsoft 365 Copilot discount below does not apply — every run is billed.

Estimate the two regimes separately

In the Credit Estimator, the Detailed and Quick + Import modes let you flag an agent as interactive or autonomous and dial in a buffer for the share of interactions that escalate or fire extra tools — so you plan for the average path and the busy one.

The Microsoft 365 Copilot license — what it covers

If your users hold a Microsoft 365 Copilot license, most employee-facing agent usage on the standard harness is zero-rated — it doesn't draw down the credit pool at all.

Per Microsoft's billing rates (re-verified 2026-10-28), every standard-harness meter — classic answers, generative answers, agent actions, tenant graph grounding, agent flow actions, and the AI tools — shows "No charge" when it's used by a Microsoft 365 Copilot-licensed user, for business-to-employee scenarios where the agent runs under that user's own authenticated identity, subject to fair-use limits.

This got broader in 2026

Earlier guidance limited the zero-rating to classic answers, generative answers, and tenant graph grounding in Copilot Chat, Teams, and SharePoint. The current rate table extends "No charge" to every standard-harness feature for licensed employees. Microsoft's licensing page and the Licensing Guide remain the final word for contracts.

The exceptions that still bill, even for licensed users:

  • GitHub Copilot harness agents — never covered by the license. Every interaction — plus building and testing the agent — bills Copilot Credits, regardless of channel or who's licensed. Microsoft bills this harness per task, by complexity, bundling model tokens, tools, and the harness itself, and shows per-task ranges only in a diagram on its billing overview. The estimator's Light 100–300 · Medium 300–500 · Heavy >500 credits per task anchors come from that diagram — treat them as illustrative. See Pick the engine.
  • Customer Conversation harness agents — the newer harness for customer-facing voice and chat (contact center–style) is billed in Copilot Credits on usage. Customer-facing use isn't employee-facing anyway.
  • Autonomous / triggered runs — no licensed user's identity in the loop.
  • Agent flows on any trigger other than "When an agent calls the flow."
  • Computer-Using Agent (CUA) actions — never included in the Microsoft 365 Copilot license.
  • Unlicensed or external users — anyone without the license pays the full rate card.

Gross vs. net — what you actually pay

Because coverage flips entirely on the harness, the same agent at the same volume can cost very different real money. On the standard harness, licensed employees are zero-rated, so net billable can be far below gross consumption — sometimes ~$0 incremental. On the GitHub Copilot and Customer Conversation harnesses, net always equals gross. The Credit Estimator now models both: pick your harness and it shows gross consumption and net billable side by side.

Where over-runs come from

Capacity is enforced monthly and unused credits don't carry over. On prepaid capacity, custom agents are disabled at 125% of capacity until it's increased or resets (a linked pay-as-you-go meter avoids that), and admins can cap individual agents in the Power Platform admin center. Surprises almost always trace back to one of four things: an autonomous agent nobody counted per-event, a flow on a non-"agent calls the flow" trigger, volume from users who aren't licensed, or a GitHub-harness agent billed from its first test run.

How you get credits

Copilot Credits are funded at the tenant level, and there are three ways to buy them. You can mix them — and when more than one is available they're drawn down in a fixed order (below). Copilot Studio, Cowork, Autopilot, agentic SharePoint, and the Work IQ APIs all meter in the same currency; Copilot Studio capacity is managed in the Power Platform admin center, while Cowork and the other Microsoft 365 experiences are governed by spending policies in the Microsoft 365 admin center (Copilot → Cost Management).

Model What it is Commitment Where you buy it Best for
Capacity packs License-based blocks of prepaid credits (25,000 credits / month per pack) Subscription Microsoft 365 Marketplace — managed & allocated in the Power Platform admin center A steady, known monthly baseline
Pre‑Purchase Plan (P3) A commitment of discounted credit units for a one‑year term 1‑year prepay Azure portal or Microsoft 365 admin center — shows as a reservation in Azure Cost Management Committed annual volume you want to discount
Pay‑as‑you‑go (PAYG) Metered billing to a linked Azure subscription — $0.01 per credit None An Azure subscription linked to the environment or billing policy Getting started, spiky usage, and covering overflow

The draw-down order — why it matters for your bill

When more than one source is available, Microsoft consumes them in a fixed order so spend stays predictable:

1. Capacity packs → 2. P3 prepaid credits → 3. Pay‑as‑you‑go.

PAYG only kicks in after your prepaid capacity (packs and P3) is exhausted — which is exactly why a small always-on PAYG link makes a useful safety net on top of prepaid capacity rather than a runaway cost.

P3 vs. packs — the quick read: both are prepaid, but capacity packs are a license subscription you buy in the M365 Marketplace and allocate per environment, while a Pre‑Purchase Plan (P3) is an Azure reservation — a one‑year commitment that trades flexibility for a discount on the credit rate. PAYG is the no‑commitment meter that backs both and catches the overflow.

For the dollar conversion behind these, use the Credit Estimator (it shows both pay‑as‑you‑go and prepaid pricing on every estimate) or the Copilot Studio Licensing Guide for the authoritative rates.

Put a number on your agent

  1. Describe it or upload it in the Credit Estimator. Quick mode takes a plain-English description; Detailed models it feature-by-feature; the Solution-package mode analyzes a real exported agent.
  2. Read the two axes — Build effort (the T-shirt size) and Run cost (credits/month) — as separate answers.
  3. Sanity-check the regime — interactive or autonomous — and whether your users are licensed, because that's what decides how much of the rate card you actually pay.

Estimating high and the number scares you? That's often a right-sizing signal — see When you've gone too far before you build.

Where this leads

Sources

All product and rate claims on this page are grounded in official Microsoft documentation, last re-verified 2026-10-28:

Claim Source Retrieved
Copilot Credits are the metering unit; full standard-harness rate card (1 / 2 / 5 / 10 credits, flows 13/100, AI tools, 8/page, voice 10/35/75 per min); 10+2=12 stacking example; reasoning-model dual meter at 10 credits / 1K tokens, prompt-cache reads unbilled; "No charge" for every feature for M365 Copilot-licensed B2E users (CUA excluded; flows only via "When an agent calls the flow"); 125% overage enforcement Billing rates — Copilot Studio 2026-10-28
Everyday AI (USL) vs. advanced AI (usage-based billing: Cowork, Code, Autopilot, frontier models); UBB requires a USL and an admin billing policy USL and usage-based billing 2026-10-28
Three harnesses (GitHub Copilot, Customer Conversation, standard); GitHub Copilot harness billed per task incl. build & test, never license-covered Harnesses overview · Billing for the GitHub Copilot harness 2026-10-28
"messages → Copilot Credits" rename (Sept 1, 2025); credits pooled per tenant; pay-as-you-go / prepurchase / prepaid packs; monthly enforcement, no carryover; Microsoft 365 Copilot zero-rating in Chat/Teams/SharePoint Standard harness licensing — Copilot Studio 2026-08-04
Three purchasing plans — Capacity packs, Pre‑Purchase Plan (P3), Pay‑as‑you‑go; P3 is a one‑year discounted reservation; capacity packs bought in the M365 Marketplace & allocated in the Power Platform admin center; fixed draw‑down order packs → P3 → PAYG Manage usage-based billing — M365 Copilot · Capacity packs · Pre‑Purchase Plan (P3) 2026-08-19

Unofficial and community-built. This page is not endorsed by or affiliated with Microsoft. It summarizes public documentation to help you learn — always confirm customer-facing numbers against the Microsoft sources linked above.

Was this page helpful?

Found an issue or want to suggest an improvement? Open a GitHub issue  ·  Star on GitHub