M365 Copilot Agents — Cost Model Reference
A clear, sourced breakdown of how Microsoft charges for AI agents in M365 — what is free, what triggers billing, and how to avoid the three most common budget surprises. Includes five concrete scenarios with math.
The main trap: many organizations assume their M365 Copilot license covers all agent usage. It covers a lot — but not everything. Autonomous agents, external users, and heavy usage can all generate surprise charges. This reference explains exactly where the line is.
New as of the June 16, 2026 GA wave: Copilot Cowork and the Work IQ APIs both reached GA — and both meter through Copilot Credits on top of the flat per-seat licences. The $30/user/month Copilot add-on is unchanged; the budget now has a fixed line (seats) and a variable line (Cowork, Work IQ grounding, and agent runtime) that behaves like cloud spend.
Microsoft's multi-step agentic execution layer (built with Anthropic) moved from the Frontier preview to GA worldwide on June 16, 2026. It requires an M365 Copilot licence and is off by default, but the work it does is billed separately on usage in Copilot Credits ($0.01/credit PAYG). A single Cowork task's cost stacks four inputs:
The Work IQ APIs reached GA on June 16, 2026 — and this is the distinction that matters for anyone building agents:
Sources: Copilot Cowork is now GA (Jun 16, 2026), the new Work IQ APIs (Jun 2, 2026), usage-based billing (Copilot Credits).
| Agent activity | Licensed employee | External / unlicensed | Autonomous trigger |
|---|---|---|---|
| Scripted / FAQ answer | Free | 1 credit | 1 credit |
| AI-generated answer | Free | 2 credits | 2 credits |
| Search company data (SharePoint, Teams) | Free | 10 credits | 10 credits |
| Agent action (reasoning, workflow step) | Free | 5 credits | 5 credits |
| Autonomous trigger (agent acts without being asked) | ⚠ 5 credits — never zero-rated | ⚠ 5 credits — never zero-rated | ⚠ 5 credits — never zero-rated |
“Free” = zero-rated under fair usage for M365 Copilot licensed users. Microsoft's fair usage threshold is not published. 1 prepaid credit = $0.008. 1 PAYG credit = $0.01.
Claude Sonnet and Opus are natively selectable as the AI model inside any Copilot Studio agent (available since January 2026, admin opt-in required). Using Claude Sonnet bills at the same Copilot Credit rates as Microsoft's own models — no premium.
Anthropic publishes Claude as an Office add-in (GA May 7, 2026). This is not an M365 Copilot feature. It requires a paid Claude plan (Pro, Max, Team, or Enterprise). No Copilot Credits are consumed and it does not appear on your Microsoft or Azure invoice.
A Copilot Studio agent that calls Claude via Azure Foundry generates two separate charges: Azure Marketplace token costs (Anthropic pricing) on the Azure invoice, and Copilot Credits for agent orchestration on the Power Platform invoice. Both fire simultaneously.
Microsoft built Copilot Cowork in collaboration with Anthropic using Claude. It reached GA worldwide on June 16, 2026, moving out of the Frontier preview. It requires an M365 Copilot licence to access, but the work it does is billed separately on usage in Copilot Credits ($0.01/credit PAYG), and it is off by default. A task's cost stacks four inputs: model use + context retrieval + tool calls + runtime — the model picker (Opus 4.8 / Sonnet 4.6 today; a lower-cost "Cowork 1" coming) is the main lever on cost per task.
Need help deploying M365 Copilot agents at scale?
The M365 Copilot Deployment Kit packages the agent templates, rollout roadmap, governance checklist, and ROI math a deployment lead needs. Buy once, own it, use it that week.