Thirty dollars, and then the meter

Microsoft put its business artificial intelligence into one application on Friday. The new Copilot has three sections — Home, which folds chat, the Cowork agent and full versions of Word, Excel and PowerPoint into one surface; Code, built on GitHub Copilot's technology; and Autopilot, the renamed Scout, a cloud-hosted agent with its own identity and email address. The commercial document published beside it matters more than any of it.

That document is a post by Nicole Herskowitz titled "Evolution of the Copilot pricing model", and it splits the product in two. Everyday AI — chat, and Copilot inside the Office applications — stays on a user subscription licence, which for commercial Microsoft 365 Copilot is $30 per user per month. Advanced AI — Cowork, Code, Autopilot and the newest frontier models — runs on usage-based billing, metered in Copilot Credits. The sentence that matters to anyone modelling this is short: usage-based billing "builds on the USL, which is required to access these capabilities." The seat is not replaced. It becomes the entry ticket, with the meter on top.

Microsoft is explicit about the target. "Some vendors put everyday AI work on a meter," the post says. "We think that's the wrong deal, because when every task has a meter, companies may ration access and only some people get to use AI." Satya Nadella told investors in July that the model was moving "beyond per-seat to per-seat-plus-consumption".

Underneath the argument sits a cost comparison Microsoft ran itself, and it needs a flag: it is published only as an image, with no alternative text and no figures repeated in the prose. Read off the image, the claim is this. For a standard user running fifteen everyday tasks: Copilot with Auto, $30, all of it subscription; Claude with Opus 5, $84, being $20 of access and $64 of meter; ChatGPT with GPT-5.6 Sol, $38. For an advanced user running twenty-five tasks, five of them long-running: Copilot with Cowork on Opus 5, $73, being the $30 seat and $43 of meter; Claude with Claude Cowork on the same Opus 5, $149, being $20 and $129 of meter.

Take it at face value and the position is remarkable. Microsoft is telling corporate buyers that the identical Anthropic model, on the identical workload, costs $43 through its meter and $129 through Anthropic's. The footnote says each of forty prompts was run ten times, and that competitor costs used publicly available list pricing and API rates, excluding discounts. It is a vendor's benchmark against rivals' rack rates and should be read as one. It is also the first public statement of what a bundler thinks a frontier model is worth as a component of someone else's product.

That is the part the model vendors should be reading. Bundling does two things to them, neither about features. The first is that it decides which models sit inside the flat fee: GPT-5.6 and Sonnet 5 are included in the licence, Opus 5 "will be included with limits", and the newest models — OpenAI's Astra and Anthropic's Fable — are metered. The post goes further: "some capabilities will start on UBB and move into the USL as the economics allow." That is a graduation schedule Microsoft sets, for models it does not build, against economics only it can see.

The second is routing. Model selection is part of the licence, but the default is Auto, which "weighs accuracy, speed, and cost on each request to route to the model and level of reasoning effort best suited to the job". The justification is a concept Microsoft calls saturation: the point at which a more capable model stops changing the answer because the task never needed it. Inside a fixed price, every request is one Microsoft has a direct financial interest in routing to the cheapest adequate model. Administrators can also restrict which model families a group may use, and those settings, the post notes, "also shape which models Auto can select from".

Set that against what this desk reported on Thursday. McKinsey's survey found the unit price of intelligence falling while invoices rose, one organisation in five saying cost now limits what it does with AI, and the cost of the same task varying thirtyfold between runs. Microsoft's answer to that variance is not cheaper tokens but governance: metered services stay off until an administrator writes a spending policy, and budgets are set at tenant, group and user level. The finance function gets the dashboard before it gets a predictable number. Jacob Andreou, the executive vice-president who runs Copilot, was candid with CNBC: "This product is at its best when it's just optimizing for user value and for the quality of the output." Cost efficiency, he added, is less important. That is an honest description of a meter.

The reason any of this is urgent is adoption. Microsoft reported more than 30 million paid Copilot seats in July, up from 20 million in April, against more than 450 million commercial Microsoft 365 seats — under seven per cent. Microsoft 365 cloud revenue reached $100.3bn in fiscal 2026, up about 19 per cent, with no Copilot breakout. The shares are up three per cent this year, behind the Nasdaq and every megacap peer.

So the bundle is doing two jobs at once. It defends a per-seat business that has converted one enterprise seat in fourteen, by making the seat compulsory for anything interesting. And it turns two model vendors into components whose price, placement and traffic share are set by the company reselling them. Both spent the past fortnight cutting their own prices. On Friday's numbers, the cut lands in somebody else's margin.