Capacity pack

$200/mo per 25,000 credits

Tenant-pooled, packs stack, no rollover

Pay-as-you-go

$0.01/credit via Azure

No commitment, needs an Azure subscription

Pre-Purchase (P3)

5-20%off, 1-year upfront

Tiers from $2,850, bought in the Azure portal

The maker license (building agents in Copilot Studio) is free; admins assign it in the admin center. You pay for what agents consume, not for who builds them.

US list prices. In September 2025 Microsoft renamed "messages" to Copilot Credits with no change to pack size or the pay-as-you-go rate, so older "message pack" documentation describes the same $200/25,000 product.

Why two agents with the same traffic have wildly different bills

Every feature an agent uses draws a different number of credits: a scripted classic answer is 1, a generative answer 2, an agent action 5, tenant graph grounding 10, and premium AI tools (the reasoning-model meter) 100 per ten responses, billed at 10 credits per 1,000 tokens. Reasoning models bill on two meters at once: the feature rate plus the premium token rate. The complete rates table with planning guidance lives in the companion piece, Copilot Credits and capacity packs, explained.

Design, not traffic, is the first cost decision: grounding toggles, orchestration mode, and model tier set your per-interaction rate before the first user shows up.

Three worked scenarios

These are adapted from Microsoft's own billing examples, with the monthly math added (assumed 22 working days, pay-as-you-go rate):

1. Website support agent (external users)

Four classic answers plus two generative answers per visit, 900 visitors a day: (4×1 + 2×2) × 900 = 7,200 credits/day.

≈ 158,400 credits/month ≈ $1,584 on the meter, or 7 capacity packs ($1,400). Packs win at this volume.

2. Graph-grounded sales agent (mixed workforce)

Four generative answers plus four graph-grounded responses per run: 48 credits per user per day. 50 licensed users pay nothing (covered by their Microsoft 365 Copilot seats); the 100 unlicensed users bill (4×2 + 4×10) × 100 = 4,800 credits/day.

≈ 105,600 credits/month ≈ $1,056 on the meter. Note the shape: the license status of the audience, not the agent, drives this bill.

3. Autonomous order-processing agent

Triggered per incoming order, four action calls per run: 4×5 = 20 credits per order.

1,000 orders/month ≈ 20,000 credits ≈ $200: exactly one pack, or the meter, at break-even. Autonomous agents run under their own identity, so no seat covers them.

What to know before you commit

Packs beat the meter above 20,000 credits/month

Below that, pay-as-you-go is cheaper. Pair them and overflow bills to the meter automatically with no interruption.

Unused pack credits expire monthly

No rollover. Size packs to your floor, not your peak, and let the meter absorb the peaks.

Enforcement is real

Prepaid-only tenants get agents disabled at 125% of capacity; agent flows block new runs when capacity is fully consumed. Details and the per-agent cap that prevents both are in the credits guide.

P3 is for known volume

The Pre-Purchase Plan trades a one-year upfront commitment for 5-20% discounts. Buy it after you have real usage data, not before.

Studio credits are one slice of the wider agent bill. Seats, Agent 365 governance, and the Cowork and Work IQ consumption layer are covered in the full cost model reference; who needs which license is in the agents licensing guide.