Why Your Microsoft 365 Copilot Bill Went Variable in 2026 (and How to Control It)

For three years, a Copilot licence was a seat. You bought it per user, you deployed it, and you knew the cost on day one. That model is closing. The agentic part of Copilot now runs on a meter, and the meter has its own currency: the Copilot Credit. The gap between a controlled credit budget and a runaway one is the gap between a few hundred dollars a month and a five-figure line on the Azure invoice.

If your Copilot bill jumped this year and no one can explain it, this is why.

What changed

On September 1, 2025, Microsoft retired per-message billing for Copilot Studio and replaced it with a single metered currency. Custom agents have metered on Copilot Credits since then. Through 2026, consumption billing spread across the wider Copilot surface: Copilot Cowork and Work IQ reached general availability in mid-2026, and the grace period for Frontier-program tenants ended on July 1, 2026. Any custom agent built in Copilot Studio, Azure AI Foundry, or a third-party framework now meters Copilot Credits on every interaction. Prebuilt Microsoft 365 Copilot agents stay covered by your per-user licence. The custom agents your teams built freely are the ones running the meter.

A Copilot Credit costs one cent at pay-as-you-go, or eight tenths of a cent inside a prepaid pack. The price of a credit is not the problem. The number of credits each action burns is.

The 100x problem

Here is the rate card that most budgets never see:

Agent action	Credits	Cost at $0.01
Classic scripted answer	1	$0.01
Generative answer	2	$0.02
Agent action (tool or step call)	5	$0.05
Content processing (per page)	8	$0.08
Tenant Graph grounding (per response)	10	$0.10
Premium reasoning (per 10 responses)	100	$1.00

A scripted answer costs one credit. A reasoning-heavy response runs a hundred credits per ten calls. That is a hundred-fold spread inside the same product. Two agents can look identical to an end user, take the same volume of questions, and land bills fifty times apart, because one reasons on every turn and the other does not.

The costs also stack inside a single interaction. Microsoft’s own example: a tenant-graph-grounded agent costs twelve credits per response, ten for the grounding and two for the answer. Add two tool calls and you reach twenty-two. Route the same turn through a reasoning model and one answer passes a hundred credits. Nothing in the builder warns you as you assemble it.

The lesson lands in one line. The architecture is the budget. Credits pool at the tenant level, so cost does not track how many people use an agent. It tracks how the agent is built. Grounding, reasoning, and tool calls are the expensive verbs.

Why capping the bill does not fix it

Enterprises reach for the obvious levers first, and each one falls short:

• A spending limit caps total spend and hard-stops your agents at 125% of capacity. It does not lower the per-call grounding and reasoning charge, and a hard stop means suspended agents your business depends on.
• A prepaid pack caps the pool at a lower unit rate. It does not remove the variable charge inside each call.
• Default-denying tools controls access. It does not make a reasoning agent cheaper.

The cost conversation has left procurement and moved into engineering. The people who design the agents set the run-rate, so a cost review that never looks at agent topology reviews the wrong thing.

How to get back in control

Three moves, in order:

1. Inventory and price the estate. List every custom agent and Copilot-connected app, then model the credit burn per interaction against real volume. Most teams have never seen this number per agent.
2. Classify each agent. Keep the cheap scripted bots where they are. Flag the reasoning-heavy, high-volume agents that drive the bill. Retire the duplicates no one uses.
3. Move the expensive agents to Azure AI Foundry. Foundry-native agents bill on Azure consumption instead of the Copilot Credits pool, run with Azure Cost Management spending limits and hard caps, and give you observability on every call. For heavy custom agents, the cost model works in your favor, and the workload anchors to Azure where Microsoft funding follows.

Deciding which agents to keep and which to move is its own step. Our guide on Copilot Studio vs Azure AI Foundry walks through the keep, move, or retire call for each agent. When you are ready to move, the Copilot to Azure AI Foundry migration guide covers the mechanics and the Microsoft funding.

The point is not to abandon Copilot Studio. Keep it as the front door for lightweight, business-facing copilots. Move the agents that meter hardest to Foundry, where cost is governed and predictable.

FAQ

Why did my Copilot bill increase in 2026? Custom agents shifted from a fixed per-seat licence to consumption billing on Copilot Credits. Reasoning, grounding, and tool calls each add credits, so heavy agents cost far more than the old flat model implied.

Do prebuilt Microsoft 365 Copilot agents meter credits? No. Prebuilt M365 Copilot agents stay covered by licensing. Custom agents built in Copilot Studio, Foundry, or third-party frameworks meter Copilot Credits.

What happens if I exceed capacity? At 125% of purchased capacity, Microsoft enforces the limit and can suspend agent access. Tenants that never configured usage-based billing have seen agents cut off.

Is moving to Azure AI Foundry cheaper? For high-volume, reasoning-heavy agents, moving to Foundry usually lowers and stabilizes cost, because Foundry-native prompts and workflows carry no per-credit charge and you gain per-agent cost controls. Confirm the before-and-after per agent before you commit.

---

Get the number for your estate. Onepane runs a two-day Agent Readiness Assessment that inventories every Copilot agent, models current Copilot Credits spend against projected Azure consumption, and shows you which agents to migrate, keep, or retire. Microsoft funding covers most of the work. Model your ROI or book a demo.