Your Flat AI Subscription Is Becoming a Meter

The whole point of a subscription is that the meter is switched off. You pay a fixed amount, you use the thing as much or as little as you like, and the bill at the end of the month is the same number you agreed to at the start. That predictability is not a minor perk; for most people it is the entire reason they chose a subscription over paying by the drink. You are buying certainty as much as software.

In early August 2026, Microsoft published the documentation for a billing model that quietly reverses that deal. It is called Copilot Credits, and it sits alongside your fixed licence as a “flexible payment option aligned to actual usage” — which is a polite way of saying the meter is being switched back on. The fixed fee does not go away. It is joined by a second, variable one, denominated in a unit you cannot count in advance and will only fully understand after the invoice arrives.

This is not a Microsoft eccentricity. It is where the whole industry is heading, and it is worth understanding the shape of it before your own plan grows a meter you did not ask for.

The announcement, in Microsoft’s own words

The primary source here is not a leak or a hot take; it is Microsoft’s own Learn documentation, published in the first week of August 2026. It describes “usage-based billing” that “charges customers based on actual usage, measured in Copilot Credits,” and states that this “complements fixed subscription licensing with a flexible payment option aligned to actual usage.” Credits are positioned as “a common currency for eligible Microsoft services with usage-based billing.”

Then comes the sentence that tells you where this is going: “In some scenarios, licenses act as an entry point enabling access to AI services billed on a pay-as-you-go basis.” Read that twice. The licence — the flat fee you already pay — becomes the ticket that lets you start spending more. At launch this applies to newer agentic services such as Cowork and the Work IQ API, but Microsoft is explicit that it “will add more agents and services over time.” The direction of travel is not subtle.

To Microsoft’s credit, the same documentation describes genuine cost controls: an admin dashboard with budgets, alerts, and “hard caps” to “prevent overspending,” plus a Cowork usage estimator. We will come back to those, because they matter. But note what their existence concedes: overspending is now a thing that can happen to you. That was never true of a flat subscription. You do not need a spending cap on a fixed monthly fee.

This did not start in Redmond

If the Copilot Credits model feels familiar, that is because Microsoft’s own GitHub arm did it first. In April 2026, GitHub announced that from 1 June, all Copilot plans would move to usage-based billing. Premium request units were replaced by “GitHub AI Credits,” consumed “based on token consumption, including input, output, and cached tokens, using the listed API rates for each model.” The flat developer subscription that people understood became a token meter that most of them did not.

The rollout was not universally loved; it arrived against visible pushback from paying customers, some of whom experienced the “improvement” as a rate limit wearing a new hat. What is striking is how quickly the pattern has propagated from a developer tool that at least talks in tokens for a living to a mainstream productivity suite used by people who have never heard the word “token” and have no reason to.

Metering is not, in itself, a swindle

Here is where fairness is owed, because the companies have a real argument and it deserves to be stated at its strongest. Running these models genuinely costs money, per request, forever — a point we have made at length in our piece on token costs. Unlike a traditional software feature, which costs almost nothing to serve to one more user, an AI feature bills real compute every single time it runs.

GitHub put the case plainly in its own announcement: “a quick chat question and a multi-hour autonomous coding session can cost the user the same amount,” and absorbing that gap under a flat fee was, in its words, “no longer sustainable.” That is true, and it is not spin. Under a flat subscription, the light user quietly subsidises the person running autonomous agents around the clock. Usage-based pricing is, in the abstract, the fairer model: you pay for what you actually consume, and the heavy user finally carries their own weight. A vending machine that charged everyone the same regardless of how many drinks they took would be a strange vending machine.

So the objection is not to metering as a principle. The objection is to metering done in a currency the customer cannot read, bolted onto a fixed fee that was sold as the whole price, with the predictability quietly removed and nothing offered in its place.

A meter you cannot read is not really a price

The deepest problem is the unit. Copilot Credits, like GitHub AI Credits before them, ultimately track something close to token consumption — and, as we have argued before, nobody thinks in tokens. You think in tasks: draft this email, summarise that thread, have the agent tidy the spreadsheet. The relationship between “a task I want done” and “the number of credits it will cost” is opaque at the moment you incur it, and only reconciles after the fact, on a dashboard you have to go and look at.

Every other metered utility you live with gives you a gauge you can actually read. Electricity is kilowatt-hours; the taxi is miles and minutes; the mobile plan shows your data allowance ticking down. You can form an intuition, glance at the meter, and change your behaviour before the bill lands. An AI credit has no such intuitive anchor. The estimator tools and consumption dashboards Microsoft ships are a real attempt to help, and better than nothing — but needing a modelling tool to guess next month’s bill is itself the tell. You do not budget for a fixed subscription. You just pay it.

The flat fee becomes an entry ticket

The genuinely novel move — and the one worth watching — is that the meter does not replace the subscription. It joins it. The licence, in Microsoft’s own framing, becomes “an entry point” to services you then pay for again by consumption. You are not being moved from a flat plan to a usage plan. You are being asked to keep the flat plan and feed the meter.

This is the same pattern we described in our piece on subscription fatigue, where capabilities quietly migrate up the tiers and the floor moves beneath you. Here the mechanism is cleaner: the capability you thought was included becomes the thing you pay extra to actually use at any scale. The most useful new features — the agents that do multi-step work on your behalf — are precisely the ones that consume the most credits, which means the better the tool gets at the thing you bought it for, the more it costs you to let it do that thing. The incentive to build a concise, restrained agent is not obviously on the vendor’s side.

Microsoft is not doing this alone

Zoom out and the flat-fee era looks less like a settled market and more like a promotion that is now ending on schedule, from both directions at once. In China, DeepSeek — the company whose rock-bottom prices did more than anyone’s to anchor the market’s expectations — told developers on 6 August 2026 that its API prices would rise “significantly” in the near future, and that they should “plan their usage accordingly.” That warning came days after it shipped an ultra-cheap new model. The cheap price got the users; the users got the price rise.

Meanwhile OpenAI spent the same week removing message limits for free ChatGPT text chats — the giveaway end of exactly the same strategy. Acquire on generosity; monetise later, once the habit has formed. Free-and-unlimited at the front door and metered-and-uncapped at the back are not opposites. They are the two ends of a single funnel, and most users only ever see one end at a time.

What honest metering would look like

None of this is an argument for freezing prices or pretending compute is free. It is an argument for metering that treats the customer as someone owed information rather than a meter to be read after the fact. A version that respected the person paying would:

• Price in something a human can feel. If the unit must be credits, peg them transparently to tasks or actions, not to a token count the user has no way to predict.
• Show the meter live, before the spend. A running balance and a clear “this action will cost roughly X” prompt, the way a cash machine warns you about a fee before you accept it — not a dashboard you have to remember to check.
• Make hard caps the default, not an option. Microsoft’s caps and alerts are the best part of this design; they should be switched on out of the box, so overspending requires opting in rather than opting out.
• Keep an honest flat option. Plenty of people will happily pay a predictable premium to make the meter go away. Offering that is not a loss leader; it is selling certainty to the people who value it, which is most of them.
• Not double-charge for the same capability. If the licence is an “entry point” and the usage is metered, be plain about what the licence actually still buys, rather than letting it quietly become a cover charge.

The frustrating thing is that Microsoft has clearly built some of this. The caps, the budgets, the estimator — the machinery for a legible meter is right there in the documentation. What is missing is the disposition to point it at the customer’s peace of mind rather than the administrator’s spreadsheet.

What to do before your plan grows a meter

For now, the practical stance is the same one we recommend for any AI billing model, and it is not complicated:

• Find out whether your plan is affected. Metered credits are arriving first on the newer agentic features; check whether the things you rely on are billed by licence, by credit, or by both.
• Turn hard caps on immediately. If your provider offers a spending limit, set it low and raise it deliberately. An automated agent does not get bored, and, as we noted in the token-costs piece, the bills that shock people almost always come from automation nobody was watching.
• Estimate before you enable. Use whatever consumption estimator exists before switching on a metered feature at scale, and treat the demo’s cost as the floor, not the average.
• Watch the total, not the headline. The relevant number is licence plus consumption, not the sticker price of the seat. A cheap-looking plan with an open meter can cost more than an expensive-looking flat one.

The flat subscription was never a law of nature; it was a customer-acquisition offer, and offers end. Usage-based pricing may even be the fairer settlement in the long run. But fairness runs both ways, and the version currently shipping asks the customer to give up the one thing the subscription was for — knowing the bill in advance — while handing back a dashboard and calling it control. Predictability is a feature, much like the legible allowance we keep asking rate-limited products to show. Charge me by the drink if you must. Just let me read the meter before I’m thirsty.