--- title: "Microsoft Put a Price on Copilot Chat. The Agents Get a Meter and No Rate Card." description: "A breakdown of Microsoft's new Copilot app, why every agent feature moved to a meter with no published rate, and the spend controls I'd set before switching Autopilot on." author: "Jahanzaib Ahmed" date: 2026-09-26 category: "trends" readingTime: "13 min read" tags: ["ai news", "ai-agents", "microsoft", "ai-pricing"] canonical: https://www.jahanzaib.ai/blog/microsoft-copilot-pricing-agents-usage-billing source: https://www.jahanzaib.ai --- # Microsoft Put a Price on Copilot Chat. The Agents Get a Meter and No Rate Card. Two years ago Microsoft wanted you to buy a laptop with a 40 TOPS neural chip so AI could run on your own desk. On Friday it launched an agent that runs on its own computer in Microsoft's cloud, keeps working while you sleep, and is billed by the unit. That is the real Microsoft Copilot pricing story from September 25, and it's bigger than the three new tabs everyone wrote about. The chat part of Copilot has a price. The agent part has a meter, a set of spending controls, and no published rate. Here's what shipped, what it costs as far as anyone can tell, and what I'd set up before a single agent gets switched on. ## What did Microsoft actually launch on September 25? A single Copilot app with three tabs: Home, Code and Autopilot. Home merges the old Chat and Cowork modes into one box, Code lets anyone build small internal apps, and Autopilot is the renamed Scout agent, now running in the cloud. Home and Code go to Microsoft's Frontier early access program first, and Autopilot enters private preview at the end of the month. The details matter more than the tabs. [Microsoft's launch post](https://blogs.microsoft.com/blog/2026/09/25/introducing-the-new-copilot-with-home-code-and-autopilot/), written by Jared Spataro, describes Cowork as the mode where "you define the task and it runs end-to-end to return a completed result." It also adds Office in Copilot, so a request can produce a real Word document or Excel workbook next to the conversation instead of a paste job. It's also one app for both audiences. The interface is identical whether you sign in with a personal or a work account; a work account simply reaches company data through Microsoft's Work IQ APIs. Jacob Andreou, the Copilot executive VP, said in recorded launch remarks that about 90 million people pay for Microsoft 365 themselves, [as ZDNET reported](https://www.zdnet.com/innovation/microsoft-new-copilot-unified-ai-app/). That's Microsoft's argument for not shipping consumers a cut down version. Code is the surprise. It's built on the same technology as GitHub Copilot, runs in a sandbox, and hosts what you make inside your own Microsoft 365 tenant. The Verge's Tom Warren called it ["not one you'd typically associate with an app designed for knowledge workers"](https://www.theverge.com/news/1000532/microsoft-copilot-super-app-chat-coding-autopilot), which is fair. Microsoft's ambition isn't small either. Spataro's line, quoted by The Verge: "Just as Office defined work for the PC era, the new Copilot is built to define it for the AI era." And Autopilot is the piece I'd watch. Microsoft calls it "a persistent, proactive and personal agent that keeps working even when you're not." It watches channels, follows up on threads and runs recurring work "without waiting for a prompt." ## How does Microsoft Copilot pricing work now? There are now two billing models inside one app. A user subscription license covers chat, Copilot inside the Office apps and model selection at a fixed monthly price. Anything that behaves like an agent (Cowork, Code, Autopilot and the long running agentic features) moves to usage based billing instead. Ed Bott at ZDNET laid this split out most clearly, and the [ZDNET piece](https://www.zdnet.com/innovation/microsoft-new-copilot-unified-ai-app/) is the one to read if you only read one. Here's the split, rebuilt from ZDNET, The Verge and Microsoft's own FinOps post: | Capability | How it's billed | Where that comes from | | --- | --- | --- | | Chat | Fixed monthly license | ZDNET, The Verge | | Copilot in Word, Excel, PowerPoint, Outlook and Teams | Fixed monthly license | The Verge | | Model selection | Fixed monthly license | ZDNET | | Cowork | Usage based | ZDNET, The Verge | | Code | Usage based | ZDNET, The Verge | | Autopilot | Usage based | ZDNET, The Verge | | Copilot Managed Runtime (hosting for built apps) | Tracked in Agent 365 cost management | Microsoft FinOps post | | Copilot Studio agents | Added to Agent 365 cost management from October | Microsoft FinOps post | ![ZDNET article section headed How much will all this cost, explaining that Copilot chat uses a user subscription license while Cowork, Code and Autopilot require usage based billing](https://cdn.sanity.io/images/qajb7q5q/production/501f7bbc155a77d25f2304a60571b7e9bf7e06fe-2880x1800.png?w=1200&q=75&auto=format&fit=max) _ZDNET's pricing section names the exact dividing line: fixed price for chat and Office features, usage billing the moment the work becomes an agent._ The Verge adds one detail ZDNET didn't: usage of newer models like Astra and Fable also lands on the meter. So the model you pick changes the bill too, not just the mode. ## Why is there no price for the agents? Microsoft published the controls before the rates. Its [FinOps for AI post](https://aka.ms/finopsforai/optimization) lists spending policies, caps and reports in detail, but I couldn't find a per credit price for Cowork, Code or Autopilot in any of the three launch posts. The company's own framing was "We're evolving Copilot pricing." Bott's verdict on the big questions (what this costs and what it pays back) was blunt: "no one knows." I agree, and I'd go further. When a vendor ships a spending dashboard and a Graph API for spend policies before it ships a rate card, it's telling you where it expects the arguments to happen. Not in procurement. In the monthly bill review. That's not automatically sinister. Agent cost genuinely is hard to quote, because it depends on how long the task runs and which model does the work. I priced a real agent loop against two frontier models [when OpenAI cut GPT-6 Sol's price](https://www.jahanzaib.ai/blog/gpt-6-sol-luna-pricing-prompt-caching-agent-costs), and the cheaper per token model wasn't the cheaper one per finished task. Per token prices already mislead. A per credit price with no published conversion to tokens or minutes tells you even less. ## What happens when Copilot picks the mode for you? Right now you choose Chat or Cowork. Microsoft says that's temporary: "simply state what you need to accomplish, and Copilot will route the work to Chat, Cowork or Code." Once that ships, the router is choosing between the flat license and the meter on every request, and the person typing won't see which one they got. None of the launch coverage I read connected those two facts, and it's the part I'd flag to any finance lead. Chat sits on the fixed license. Cowork and Code sit on usage billing. A router that decides whether a request counts as chat or Cowork is, in effect, deciding whether that request costs extra. Microsoft hasn't said the router will surface that choice to the user. Anthropic made the same merge on September 16, and I argued at the time that the old Cowork tab [was quietly acting as a permission prompt](https://www.jahanzaib.ai/blog/claude-cowork-merge-agent-permission-boundary). Microsoft's version has a second consequence. The tab was also the price tag. Model choice works the same way. The Verge reports Copilot uses an automatic model picker, and Microsoft's FinOps post says admins can now restrict model families by group: "This also impacts which models Auto can select in Cowork." That's a real control, and it's the right one. It's also a default Microsoft makes for you until someone changes it. I've written about why [routing defaults are worth real money](https://www.jahanzaib.ai/blog/stripe-openrouter-acquisition-llm-routing), and this is the same lever inside an office suite. ## Why does an agent that works while you sleep change the cost math? Because a chat bill tracks what people ask, and an always on agent's bill tracks what it watches. Autopilot follows Teams channels, chases threads and reruns recurring work on its own schedule. Its usage grows with the activity around it, not with how often its owner opens the app. Microsoft says Autopilot "lives in your tenant with its own identity, memory, computer and workspace." Andreou, quoted by ZDNET, called autonomous agents like this "absolutely terrifying to an IT admin or somebody who's trying to maintain a business posture." Microsoft's answer is to treat the agent like an employee, with permissions and an audit trail. Here's the open question I'd put to a Microsoft account team this week. The FinOps controls describe caps at the tenant, group and user level. Autopilot has its own identity. So whose cap does it spend against: the person who created it, or its own? None of the launch material I read answers that. It matters, because a per user cap that an agent can sidestep by having a separate identity isn't much of a cap. I made a related argument about [why an agent needs its own nameable identity](https://www.jahanzaib.ai/blog/amazon-blocks-meta-muse-ai-agent-identity). Microsoft has given it one. Now the billing has to follow it. ## Are Microsoft's FinOps for AI controls enough? They're better than most vendors ship on day one. Usage billing is off by default, so nobody gets an agent until an admin writes a spend policy. Caps work at the tenant, group and user level with threshold alerts, costs can be charged to departmental Azure subscriptions, and end users can see their own remaining credits inside Copilot. ![Microsoft Tech Community FinOps for AI post listing four controls: usage based billed services off by default, spending controls at tenant group and user levels, departmental billing, and detailed usage reporting](https://cdn.sanity.io/images/qajb7q5q/production/de8fa6a0d7479ae3395a8ff219ebafd8e57302e3-2880x1800.png?w=1200&q=75&auto=format&fit=max) _Microsoft's own list of spend controls. Off by default is the one that matters most: the spend policy is the rollout switch._ Two more are genuinely useful. Spend policies can be managed through the Microsoft Graph API, so they can live in the same tooling as the rest of your finance automation. And credit requests from users can be redirected to a custom URL, which means they can land in an approval flow you already run. What's missing is a stated behavior at the cap. The design Microsoft describes is about asking for more: users get notified as they approach a limit, and requests route to approval. That's fine for a chat user. It's a problem for a long running agent halfway through a supplier review. Does it pause and hold its state? Stop and leave a half finished job? I'd want that answered before Autopilot goes anywhere near a real process. I built my own version of this problem at a much smaller scale, and I learned the hard way that the cap is not the design decision. The failure mode is. My backlink tracker calls a paid scraping API on a schedule. It has a $15 a month cap, a $2 a day circuit breaker and a kill file on disk, and it fails closed if it can't read its own spend ledger. It runs at about $5 a month. I wrote those numbers up when [Rippling audited its own AI bill](https://www.jahanzaib.ai/blog/rippling-ai-spend-console-token-roi) and found roughly 10 to 15% of employees drove about 60% of the total, with token spend compounding at 80% month over month. That concentration is why per user caps matter more than the tenant total. In my experience a tenant cap only tells you the month went wrong after it already has. ## Why did Microsoft drop the Copilot+ PC label the same week? Microsoft's new Surface laptops, announced the same week, don't carry Copilot+ PC branding even though they meet its bar. Surface VP Brett Ostrum told Windows Central they "are not called Copilot+ PCs." Ars Technica, which covered the change, lists several possible reasons, including Recall's security baggage and Nvidia's upcoming RTX Spark chips. I won't claim the two announcements are connected by design. But the contrast is worth sitting with. Copilot+ PCs needed an NPU rated at 40 TOPS or better, and the new Surface models ship with an 80 TOPS chip, [according to Ars Technica](https://arstechnica.com/gadgets/2026/09/microsoft-stops-insisting-you-need-a-copilot-pc/). Satya Nadella once said Copilot+ PCs made up [15 percent of premium laptop sales](https://arstechnica.com/gadgets/2026/09/microsoft-stops-insisting-you-need-a-copilot-pc/) in the United States over a holiday season. That was a pitch for AI you buy once and run locally. This week's pitch runs the other way. The agent gets its own cloud computer. ZDNET draws the sharpest contrast with Anthropic's Claude Code, which it describes as running locally in containers or virtual machines, while Microsoft's Code runs in a sandbox hosted inside the tenant. Anthropic's own product page puts it plainly: Claude Code ["runs locally in your terminal and talks directly to model APIs."](https://www.anthropic.com/claude-code) For regulated companies that's a real advantage, because the data doesn't leave the managed cloud. It's also compute you pay for every time it runs, instead of once at purchase. ## What should IT and finance do before switching agents on? Write the spend policy first, because under this model the policy is the rollout. Then cap per user and per group rather than just per tenant, and decide in advance what an agent does when it hits a limit. Everything else is tuning. ![Microsoft Copilot Blog post titled Build where you want, run with confidence, announcing Copilot Managed Runtime, with a diagram of the Managed Runtime SDK connecting to the Managed Runtime Host](https://cdn.sanity.io/images/qajb7q5q/production/a442764432f43497e894413f9b8bb0f2c306f7ed-2880x1800.png?w=1200&q=75&auto=format&fit=max) _Copilot Managed Runtime can host apps built in Cowork, Code and Copilot Studio inside the tenant, which turns each small app into something that runs, and bills, after its builder moves on._ Here's the order I'd do it in, as someone who has shipped agents that spend money unattended: 1. **Get the rate in writing.** Ask the account team for the credit price for Cowork, Code and Autopilot, and how a credit converts to model usage. If they can't give it, run the pilot with caps low enough that not knowing is survivable. 2. **Put every Autopilot instance in its own group with its own cap.** Don't let an agent with a separate identity spend against a human's allowance by accident, and don't assume the reverse either until Microsoft documents it. 3. **Restrict Auto's model families for Cowork.** Start with the cheaper families and widen access per group once you've seen real usage. 4. **Decide the failure behavior.** For anything a long running agent touches, write down whether hitting the cap should pause the job or stop it, and test that it actually does. 5. **Watch the app inventory, not just the bill.** Apps hosted through [Copilot Managed Runtime](https://aka.ms/CopilotManagedRuntime-Blog), whether built in Code, Cowork or Copilot Studio, show up in one inventory in the Microsoft 365 admin center, and Microsoft says even an app made with Lovable can run inside the tenant through its SDK. Apps outlive their builders. Review them quarterly. If the account team can't tell you what a credit buys, the cap you set is your price. Set it like one. If you're earlier than that and still working out which processes are worth handing to an agent at all, the [AI readiness assessment](https://www.jahanzaib.ai/ai-readiness) is a quicker place to start than a Frontier pilot. For the engineering side of running agents with real budgets, I keep the long version in my [guide to agents in production](https://www.jahanzaib.ai/blog/ai-agents-production). ## Frequently asked questions ### Is Microsoft Copilot chat still a flat monthly price? Yes. Chat, Copilot inside Word, Excel, PowerPoint, Outlook and Teams, and model selection all sit under the user subscription license at a fixed monthly price. Usage billing applies to Cowork, Code, Autopilot and long running agentic features. ### How much does Copilot Autopilot cost? Microsoft hasn't published a per credit rate. Autopilot is billed by usage, and ZDNET's summary of the cost question was that no one knows yet. Autopilot is entering private preview at the end of September, so expect pricing detail to arrive through account teams first. ### Will my users get agent features automatically? No. Microsoft ships usage billing switched off. An admin has to create a spending policy that grants access before anyone can use Cowork, Code or Autopilot on the usage meter. ### What is Copilot Code? Code is a tab in the new Copilot app that builds small apps, trackers, dashboards and automations from a plain language description. It runs on the same technology as GitHub Copilot, works in a sandbox and hosts what it builds inside your Microsoft 365 tenant through Copilot Managed Runtime. ### What was Scout, and how is Autopilot different? Scout was the personal AI assistant Microsoft showed at Build, first as a desktop app. Autopilot is its renamed cloud version. It has its own identity, memory and computer in your tenant, so it keeps running when your laptop is closed. ### Can admins control which AI models Copilot uses? Partly. Microsoft's FinOps for AI update lets admins restrict model family access by group, which also limits which models Copilot's automatic picker can choose in Cowork. It's rolling out for Cowork first, with other workloads planned. > **Sources:** [Official Microsoft Blog, Jared Spataro (Sep 25, 2026)](https://blogs.microsoft.com/blog/2026/09/25/introducing-the-new-copilot-with-home-code-and-autopilot/) · [Microsoft Tech Community FinOps for AI Blog, Ryan Jones (Sep 25, 2026)](https://aka.ms/finopsforai/optimization) · [Microsoft Copilot Blog, David Blyth (Sep 25, 2026)](https://aka.ms/CopilotManagedRuntime-Blog) · [ZDNET, Ed Bott (Sep 25, 2026)](https://www.zdnet.com/innovation/microsoft-new-copilot-unified-ai-app/) · [The Verge, Tom Warren (Sep 25, 2026)](https://www.theverge.com/news/1000532/microsoft-copilot-super-app-chat-coding-autopilot) · [Ars Technica, Scharon Harding (Sep 25, 2026)](https://arstechnica.com/gadgets/2026/09/microsoft-stops-insisting-you-need-a-copilot-pc/). ## Related - [Microsoft Wrote Down What Its Models Must Never Do. No Microsoft Model Is Trained on It Yet.](https://www.jahanzaib.ai/blog/microsoft-humanist-ai-code-of-conduct) - [GPT-6 Sol Costs Half of Opus 5.5 Per Token. On an Agent Loop, It Doesn't.](https://www.jahanzaib.ai/blog/gpt-6-sol-luna-pricing-prompt-caching-agent-costs) - [Apple's Reference Image Proves a Camera Was There. It Can't Prove a Photo Isn't AI.](https://www.jahanzaib.ai/blog/apple-reference-image-ai-image-provenance) --- Canonical HTML version: https://www.jahanzaib.ai/blog/microsoft-copilot-pricing-agents-usage-billing