What a Copilot Credit is
Copilot Credits are the unit Microsoft bills Copilot Studio agent usage in. On 1 September 2025 Microsoft renamed 'messages' to 'Copilot Credits' — pack size and pay-as-you-go rate were unchanged, but the name change reflects the real model: credits scale with the work an agent does, not with conversational turns. One user question can consume a dozen credits if the agent grounds on tenant data and takes actions. Budgeting per 'message' will undercount; budget per feature.
What consumes credits
The published rates (checked August 2026): a classic (authored) answer costs 1 credit; a generative answer 2; an agent action — triggers, deep reasoning, topic transitions, computer use — 5; tenant graph grounding 10; agent flow actions 13 per 100 actions; and AI prompt tools from 1 to 100 credits per 10 responses depending on model tier, with reasoning models billed on top at 10 credits per 1,000 tokens. Rates stack: a single grounded generative answer that fires one action is 17 credits. Building and testing meter too — agent actions on the activity map during test sessions are billed like production traffic (agent flow test runs are the one exemption) — so a busy authoring team consumes real credits before launch.
Pack versus pay-as-you-go
Two ways to pay, no feature difference between them. The prepaid pack is US$200 per month for 25,000 credits, pooled across the tenant, use-it-or-lose-it monthly. Pay-as-you-go bills ~US$0.01 per credit to a linked Azure subscription at month end, with no upfront commitment (both checked August 2026; billed in USD — model your AUD exposure at the prevailing rate). Enforcement is real: on prepaid capacity, agents are disabled once consumption hits 125 per cent of what you bought, so set consumption limits per agent in the Power Platform admin centre before an enthusiastic pilot takes production down.
The Microsoft 365 Copilot zero-rating
Employee-facing agent usage is included in the Microsoft 365 Copilot licence: when a user licensed for M365 Copilot uses an agent under their own identity, classic answers, generative answers, tenant graph grounding and agent-called flows are all zero-rated, subject to fair use. This flips the modelling. Internal agents serving Copilot-licensed staff cost roughly nothing to run; customer-facing agents and agents serving unlicensed staff burn credits on every interaction. Count your Copilot seats before you size a credit pack — many Australian tenants need a far smaller pack than the raw usage estimate suggests.