Will premium models in Copilot Chat run up a bill we didn’t plan for?

That’s a fair worry, and you’re not alone. The public threads this autumn are full of people with a paid Microsoft 365 Copilot seat who chose an advanced model, ran out of “priority access” for the week, and had no way to see it coming. Adding a paid tier on top sounds like a recipe for confusion and a surprise bill.

Here’s what’s going on. Roadmap 571400, Chat: Access to UBB models via the model selector (In development, general availability targeted for October 2026; verify for your tenant), says users will be able to pick models that need usage-based billing from the Chat model selector. Crucially, tenant admins have to allocate credits before anyone can use them, and when a user runs out they can request more from admins. So it isn’t open-ended by default: no allocation, no access.

The machinery already exists in Learn. Copilot, then Cost management in the Microsoft 365 admin center is where you enable usage-based billing, create spending policies scoped to Entra security groups, choose which agents and services each policy covers, set organisation-level and per-user limits with alerts, and pick the billing method (prepaid capacity packs first, then a Copilot Credit Pre-purchase plan, then pay-as-you-go). The Credit requests page shows who is asking, their licence and current policy, and lets you add them to a policy or raise a limit. You can also set a custom request policy that redirects requests to your own IT portal or approval flow. Two settings deserve a hard look: Auto-apply new services is on by default in spending policies, so a broad existing policy could quietly cover new metered experiences; and model controls (model profiles) currently apply only to Copilot Cowork, so don’t assume you can pick which Chat models are allowed until Learn says so.

Despite the marketing, the most expensive model isn’t automatically the best one for the job. For most everyday drafting, Auto is designed to choose sensibly. Steady before scale puts the Cost & Value pane in front of the model selector: decide who has a real reason to pay for heavier reasoning before the button appears.

Options that actually work:

1. Named-group policy with a modest per-user cap. Create a spending policy for a small group (analysts, engineers, legal) with a monthly limit and alerts, turn off Auto-apply on it, and route credit requests to a business owner. Best Steady default.
2. Request-led access. Allocate nothing broadly; let the request flow tell you who actually needs it, and approve with a reason. Slower, but it gives you evidence for the business case.
3. One broad all-users policy with Auto-apply left on. Easy to set up, and the first time you’ll understand the cost is the Azure invoice.
4. Block it entirely forever. Cheap, but the people with genuine heavy work will find unsanctioned tools instead.

12 Panes action / comms: Tell users: Use Auto first; the premium models cost real money and you may need to ask. Tell FinOps and the AI administrator: Policies by group, per-user caps, Auto-apply reviewed, requests routed to someone who owns the budget. Tell the service desk: “You’ve used your priority access” isn’t an outage; here’s what it means. Leadership: Premium reasoning is a funded choice for named roles, not a default.

Benefits / risks: People who need heavier models can get them legitimately; admins see usage by user, group and service; dashboards refresh every few hours, so they’re not live; priority-access limits on included models and credit-billed models are different things and will confuse people unless you explain them. Verify Learn and Message Centre as 571400 lands.

I’d suggest reviewing your existing spending policies today, switching off Auto-apply where it’s too broad, and deciding which group gets the first allocation before the selector change reaches your tenant. You’ll be ready for the requests rather than surprised by them.

DIGITAL SKILL WORKS LTD / 12 PANES · You ask, we answer