Copilot & Foundry Agent Cost Calculator

1. Agent and billing scope

2. Forecast period and workload

3. Licence eligibility, not a blanket discount

The covered usage is employee-facing in eligible Microsoft channels, under the authenticated identity of a licensed Microsoft Copilot user.

Enter the share of usage, not headcount. Fair-use limits apply. Coverage is limited to eligible Standard / Copilot Chat usage. Work IQ APIs, computer use, external services and non-agent-triggered agent flows remain separately billable. A Microsoft Copilot licence does not cover GitHub Copilot harness runtime.

4. Billable operations in one conversation or run

Count operations, not user turns: one turn may use several meters. Include only actual generative responses. No answer or action is automatically added for a knowledge source or generic tool call.

Agent Builder: generative responses without tenant graph grounding are exempt; graph-grounded responses use the normal generative + graph meters. SharePoint agents are not automatically assigned the Agent Builder exemption.

4. GitHub Copilot harness: measured task budget

There is no universal per-answer tariff here. Supply observed or explicitly assumed all-in credits per task (including models, runtime, context and tools), plus a separate creation/evaluation budget for the selected period. Do not add already-included Work IQ consumption twice. Blank measurements block the estimate; zero is an explicit assumption, not an inferred free tier.

GitHub Copilot harness usage-based billing

4. Microsoft Foundry model, tokens and tools

Model calls and context assumptions

Each model call includes instructions, schemas, the current user message, retrieval and tool results. Additional calls in a turn include earlier intermediate outputs/tool results. Only the final visible response is carried to later turns; reasoning output is billed but not resent. This is an explicit approximation, not a provider token trace. Use measured averages, including retry calls.

Built-in tools

The file-search call fee applies to Responses API usage; storage is separate. A Code Interpreter session is not a tool call: reuse and concurrent conversations affect session counts (one-hour default lifetime; idle timeout also applies). Tool price edits are user assumptions, not verified rates.

Infrastructure for the selected period

Enter actual billable provisioned/idle hours, not just request time. Storage is billed on GB-days after the entered free allowance. Include Azure AI Search or other retrieval infrastructure in the search allowance, without double counting vector storage. For a test budget these are test-period amounts; for production they are monthly amounts. Zero means not estimated.

5. Separately billed integrations

Work IQ Tools API: 0.1 Copilot Credit per call. Work IQ Chat/Context API: variable consumption; enter measured credits rather than assuming a fixed rate. Both remain billable for licensed Copilot users, including when called from Foundry. This does not change the distinct licensing of older Retrieval/Chat APIs or native Work IQ within Copilot.

6. Purchase commitment and available capacity

Purchase scenarios assume a new, unused plan for this workload only. Credit packs have monthly capacity, annual billing and no monthly rollover. P3 pools expire after one year. Discounted allocation is not the invoice or a promise of savings. Discounts do not stack; reservations apply before P3, and Copilot Credit P3 before Agent P3.

PAYG overflow is actually configured for this environment.

Capacity monitoring is independent of the hypothetical purchase comparison. Enter the full applicable monthly allocation (not the remaining balance), then consumption already charged against it. This is a projection, not a live tenant report. Existing P3 balances and other workloads drawing from shared plans are not simulated.