What a Copilot Credit is
A Copilot Credit is the unit Microsoft uses to measure agent usage across Copilot Studio and the pay-as-you-go surfaces of Microsoft 365 Copilot. On September 1, 2025, Microsoft renamed the old "messages" currency to Copilot Credits. The quantities did not change: a prepaid pack still holds 25,000 units and the pay-as-you-go rate is still one cent per unit.
The word "credit" hides the important part: credits are consumed per feature, not per conversation. One user question can burn 1 credit or over 100 depending on how the agent answering it was built.
What consumes credits
Microsoft's current billing rates, per the Copilot Studio billing documentation:
| Agent feature | Billing rate | Used by an M365 Copilot licensed user |
|---|---|---|
| Classic answer (scripted response) | 1 credit | No charge |
| Generative answer | 2 credits | No charge |
| Agent action (triggers, deep reasoning, topic transitions) | 5 credits | No charge |
| Tenant graph grounding for messages | 10 credits | No charge |
| Agent flow actions (per 100 actions) | 13 credits | No charge |
| Text and generative AI tools, basic (per 10 responses) | 1 credit | No charge |
| Text and generative AI tools, standard (per 10 responses) | 15 credits | No charge |
| Text and generative AI tools, premium (per 1,000 tokens) | 10 credits | No charge |
| Content processing tools (per page) | 8 credits | No charge |
Two rates deserve special attention. Tenant graph grounding costs 10 credits, five times a generative answer, so an agent that reads your Microsoft Graph pays a premium on every grounded response. And reasoning models bill on two meters at once: the feature rate for the action, plus the premium AI tools rate for the reasoning itself. That is how the same agent design can cost 2 credits per answer with a standard model and far more with a reasoning model.
Read the units in that table carefully, because the premium row does not work like the two above it. Basic and standard carry two interchangeable units, a per-response rate and a per-token rate. Premium carries only the token unit: 10 credits per 1,000 tokens. Microsoft's footnote is explicit that the charge "is calculated from the exact number of billed tokens, not from the number of responses" and that "one response can involve many thousands of tokens." So there is no honest per-response figure to quote for the premium meter, and any planning model that invents one will understate it.
Microsoft's own worked example makes the size of that gap concrete. A single follow-up question to an agent orchestrated by a reasoning model bills 14,100 tokens across two model calls, which is 141 credits for one message, plus the feature rates for the activities in the turn. Tokens that the model reads back from its prompt cache are not billed. Long instructions, wide knowledge sources, and long-running conversations all push that number up, which makes instruction length and conversation hygiene a budget lever rather than a style preference.
The third column is the one most cost models leave out, and leaving it out overstates the bill. When an agent is used by an employee who holds a Microsoft 365 Copilot license, and the agent runs under that authenticated user's identity, every rate in the table is included in the license under fair-use limits. Two carve-outs: Computer-Using Agents are not included, and for agent flows the inclusion covers only runs started by the "When an agent calls the flow" trigger, so a flow on any other trigger meters at the standard rate. The meter is for everyone else: unlicensed employees, external users, and agents that run autonomously under their own identity. The full picture is in the agents licensing guide.
The three ways to pay
| Option | Price | Effective rate | Fits |
|---|---|---|---|
| Capacity packs | $200/month per 25,000 credits | $0.008 | Predictable volume; pooled tenant-wide; packs stack |
| Pay-as-you-go | $0.01 per credit, billed via Azure | $0.01 | Unpredictable or low volume; no commitment; needs an Azure subscription |
| Pre-Purchase Plan (P3) | 1-year upfront, tiers from $2,850 | 5-20% off | Known annual volume you can commit to; bought in the Azure portal |
The break-even is simple arithmetic: a $200 pack beats pay-as-you-go once you use more than 20,000 credits a month. Below that, the meter is cheaper. Above it, packs win, and the Pre-Purchase Plan wins again once you have a year of usage data you trust enough to commit against. Two behaviors matter for planning: unused pack credits do not roll over at month end, and if you pair packs with pay-as-you-go, overflow beyond the prepaid amount bills to the meter automatically with no service interruption.
The 125% rule: what happens when you run out
Without pay-as-you-go as a safety net, Microsoft allows overage consumption as a grace buffer, then enforces: when a tenant reaches 125% of prepaid capacity, custom agents are disabled. An ongoing conversation is allowed to finish; every attempt after that is rejected until you add capacity, reallocate it, or enable the meter. Your admin gets an email and a notice in the Power Platform admin center, and your users get "This agent is currently unavailable."
Agent flows enforce differently and earlier: when prepaid capacity is fully consumed (100%, not 125%), new flow runs are blocked while the parent agent keeps answering normally. In-progress runs complete, test runs are exempt, and the block resets when credits renew monthly.
The unglamorous control that prevents both: the Power Platform admin center lets you cap monthly consumption per agent (Licensing → Copilot Studio → Manage Agents) and allocate credits per environment. A cap on every production agent turns a potential outage into a visible budget line.
Five planning moves
For where credits fit in the wider agent bill (seats, Agent 365, Cowork and Work IQ metering), see the full M365 Copilot agents cost model reference. For Studio-specific pricing scenarios with worked math, see Copilot Studio pricing.