Copilot Credits Went Live: What Work IQ and Cowork Actually Cost

On June 16, 2026, two Microsoft announcements landed on the same day, and most coverage treated them as separate stories. Work IQ reached general availability with consumption billing, and Copilot Cowork became generally available. They are not separate stories. They are the same event: Microsoft turned on a single metered currency, Copilot Credits, across the Copilot Studio, Work IQ, and Cowork surface, and the meter started running the moment the announcements went live.

If you run Copilot Studio agents or have Frontier participants who can switch on Cowork, you now have a consumption line item that did not exist on June 15. This article is the cost decode: what the meter actually charges, where the real money is (it is not where the headline number points), and the governance switches an architect should set before the grace period closes on July 1. One calibration up front: the billing mechanics below are documented and live, but the day-one reliability of the underlying grounding and the quality of Cowork’s output have no production track record yet. Treat the architecture as confirmed and the maturity as unproven.

One currency, three surfaces

Until this month, the agent-cost conversation was fragmented. Copilot Studio had message packs. Work IQ was preview, unbilled. Cowork was a Frontier experiment. As of June 16, Microsoft’s licensing newsroom describes Copilot Credits as “a unified consumption currency that also covers Copilot Studio and other Microsoft AI services.” The Work IQ MCP documentation uses nearly identical language: Copilot Credits are “the common currency across Copilot Studio capabilities and Work IQ protocols across your tenant.”

That consolidation is the actual news. One pool, one pay-as-you-go rate of $0.01 per credit, one admin dashboard, drawn down by every agent capability you switch on. The two products that went GA this week are the first two taps on that pool.

Work IQ: the two-part meter

Work IQ is the intelligence layer that grounds Microsoft 365 Copilot and custom agents in organizational context. Per the Work IQ MCP overview, it is built on three layers - Data, Memory, and Inference - and exposed to agents in Copilot Studio as MCP servers that return real-time context from email, calendars, and chats. The servers available today:

• Work IQ Mail
• Work IQ Calendar
• Work IQ Teams
• Custom servers you publish

The billing has two components, and the difference between them is the whole story. From Microsoft’s licensing page, Work IQ API charges consist of:

• A static component: “0.1 Copilot Credits per API call” for actions and tools.
• A variable component for queries: grounding, retrieval, and reasoning.

At $0.01 per credit, the static charge is $0.001 per call. That is the number that traveled fastest, and it makes the feature sound nearly free. It is also almost irrelevant. The variable component is unbounded by comparison, and Microsoft’s own consumption scenarios show where the real money sits.

Now do the arithmetic. The static charge is $0.001 per call. The Light scenario bills $0.20 to $0.40 per call. That is 200 to 400 times the static charge. The Heavy scenario, at up to $1.50, is 1500 times the static charge. The 0.1-credits-per-call headline describes a rounding error sitting on top of a grounding-and-reasoning bill that is two to three orders of magnitude larger.

This is the architectural reality the per-call rate hides: an MCP server rarely makes one call. A single agent turn that reads mail, checks a calendar, and reasons over both fans out into multiple Work IQ calls, each carrying its own variable charge. The meter does not tick once per user request. It ticks once per tool invocation, and agents are built to invoke tools liberally.

Cowork: a different product on the same meter

Copilot Cowork is the other June 16 GA: an agentic system that executes long-running, multi-tool tasks end to end and is positioned to return finished work rather than drafts, though that output still needs human verification. What it brings:

• Models: Anthropic Opus 4.8 and Sonnet 4.6 today, with more coming.
• Plugins: partner plugins from firms like Harvey, Moody’s, and S&P Global.

For cost purposes, the facts that matter:

• License: requires a Microsoft 365 Copilot User Subscription License.
• Meter: billed through Copilot Credits at the same $0.01 per credit pay-as-you-go rate.
• Cost drivers: model use, context retrieval, tool calls, and runtime.
• Default state: off by default.
• Grace period: Frontier-participant organizations are not billed until July 1, 2026.

So Cowork and Work IQ are different products with different jobs, drawing from the same credit pool, gated by the same license, and governed by the same admin dashboard. If you reason about them separately you will miss the combined burn rate on a shared budget.

How much does Work IQ actually cost per call?

Microsoft’s published scenarios run $0.20 to $1.50 per call, not the $0.001 static rate: a Light task costs $0.20 to $0.40, a Medium task $0.30 to $0.75, and a Heavy task $0.50 to $1.50. The static 0.1-credit charge is a rounding error. The variable grounding-and-reasoning charge is the bill.

Here is a worked example using only Microsoft’s published numbers. Using illustrative volume assumptions - a 200-person team where each person triggers ten Medium-scenario interactions per working day, across 21 working days - at the midpoint of the Medium range ($0.525 per call):

• Per person per day: 10 calls at $0.525 = $5.25
• Per person per month (21 working days): $110.25
• Per 200-person team per month: $22,050

That is one team, one moderate workflow, before Cowork runtime, before any Heavy scenarios, before the multi-call fan-out that turns one user request into several billed calls. The pay-as-you-go meter has no ceiling of its own. The only ceiling is the spending limit you set.

The governance surface is the actual product feature

The good news is that Microsoft shipped real controls alongside the meter, and they are where an architect should spend the first hour. Per the Work IQ MCP documentation and the Work IQ APIs announcement:

• Spending limits can be set per tenant, per group, and per user across agents and services, from the Microsoft 365 admin center cost-management dashboard. This is the hard ceiling the pay-as-you-go model otherwise lacks.
• Server-level allow and block is tenant-wide. If an admin blocks a Work IQ MCP server, it blocks access for every user and every agent. “Permissions always take precedence over configuration.”
• Pay-as-you-go versus prepurchase is an admin choice. Prepurchase plans exist for organizations that want a committed, capped pool rather than an open meter.
• Observability runs through Microsoft Defender Advanced Hunting, where every tool call can be traced: which tool, which parameters, which outcome. That is your audit trail and your cost-attribution data in one place.

A meter on plumbing that is still settling

The billing is live and precise. The integrations it bills for are not always. Power Platform MVP Jukka Niiranen, testing Copilot Cowork in the Frontier preview, reported the agent stumbling on real systems:

• Dataverse MCP server tools that had been deprecated without Cowork being told.
• A Dynamics 365 Sales plugin that appeared in the configuration UI but was not actually available to the system.

His framing is that MCP is sold as the USB for AI but behaves more like USB-C, with so many variations that you cannot be sure a given combination will work until you try it. When it works, he notes, it is genuinely impressive.

That is the maturity caveat the billing announcement does not carry. As of GA you are metered per tool call on integrations that can silently deprecate or fail to resolve. The cost is deterministic from day one. The reliability is not. Budget for both, and do not assume that paying for a tool call means the tool was actually there.

Where this fits the larger Copilot Credits economy

This week is the opening of a longer arc. Copilot Studio’s own agent metering moves onto the same credit currency, which means the questions buyers ask about Studio pricing, capacity packs, and autonomous-agent consumption are now the same questions you ask about Work IQ and Cowork. One budget, one set of guardrails, one cost-attribution model across the Copilot Studio, Work IQ, and Cowork surface.

The teams that will be calm in three months are the ones that treated June 16 as a governance event, not a feature launch. The meter is on. The license gate is real. The per-call rate is a decoy. Set the spending limits, default-deny the servers, and forecast on the scenario ranges, not the headline.

What this does not mean

Five bounds so the takeaways do not over-reach.

• Copilot Credits do not replace the per-seat Microsoft 365 Copilot license. Work IQ MCP requires that license, and consumption bills on top of it.
• This meter is not the whole Microsoft AI stack. Azure AI Foundry agents bill on Azure consumption, and GitHub Copilot is a separate license and meter. The Copilot Credits pool covers Copilot Studio, Work IQ, and Cowork.
• A spending limit caps total spend and hard-stops agents when reached. It does not lower the per-call grounding-and-reasoning charge.
• Default-denying servers controls access and cost. It does not make the agents more reliable.
• Prepurchase caps the pool. It does not remove the variable grounding charge inside each call.

Read next

• Claude on Azure: The Marketplace Billing Trap - the same metered-by-default cost pattern, with named founder receipts.
• Foundry Hosted vs In-Process vs Copilot Studio Agents - which build path you are metering in the first place.
• Copilot ROI Calculator for Microsoft Enterprises - put the consumption numbers into a defensible business case.