The metering era — and the flat-rate comeback nobody’s pricing in

THE AI NEWSLETTER THAT TURNS NEWS INTO ACTION

DEEP DIVE · COMPANION TO ISSUE #138

The metering era — and the flat-rate comeback nobody’s pricing in

This week AI stopped being free to try. The major tools are converging on pay-per-token billing, and the bills are landing. But when a market races to the meter, it leaves something behind — a price you could predict. That gap is the opening, and one company is already positioned to fill it. Here’s who’s most likely to sell predictability back to you, and why a bill you can predict becomes the next premium product.

BY R. LAURITSEN · 3 JUNE 2026 · 7 MIN READ

A developer ran a single AI coding session on Monday morning and watched it eat 8% of his monthly budget in two hours. He posted the screenshot, the replies piled in, and by lunchtime the same story was everywhere: bills jumping tenfold, the old “unlimited” plans gone, people pricing out which tool to abandon. It was the day the cost of AI stopped being someone else’s problem and became a line on your invoice.

That was the visible half of the week. The same morning, GitHub moved all 4.7 million paid Copilot subscribers onto pay-per-token billing. The next day, at Build, Microsoft shipped its own models to cut what that work costs it to run. Two stories, one direction: the free-trial era is ending, and the major players are converging on the same answer — the meter.

Which is exactly why the more interesting question isn’t “how do I cut my bill?” It’s “what happens when every competitor charges the same anxious way?” Because that’s the condition under which someone breaks ranks — and the break is already visible if you know where to look.

How the market arrived at the meter

Rewind eighteen months. In 2025, AI tools competed on a single axis: capability. Whoever had the best demo that week won the early adopters, and pricing was a flat monthly fee deliberately set below cost. That wasn’t generosity. It was customer acquisition — a subsidy designed to get you hooked on a workflow before anyone worked out what it actually cost to run.

Then agents arrived, and agents burn tokens like water. A single autonomous coding session can consume more compute than a power user’s entire month of chat. Internal numbers reported by journalist Ed Zitron suggest GitHub’s weekly cost of running Copilot had nearly doubled since January. Flat-rate pricing wasn’t a pricing strategy any more; it was a bleeding wound. So the industry reached for the obvious fix: make you pay for what you use.

By June 2026, the direction is unmistakable. GitHub Copilot meters tokens. Cursor moved from flat requests to usage credits last year and ate a wave of backlash for it. Even Anthropic’s and OpenAI’s consumer plans now carry opt-in overage past the included cap. The market leaned hard toward the meter — and in doing so, handed away the one thing customers quietly valued most about the flat-rate era: a bill they could predict.

Why a predictable bill becomes the premium product

Here is the mistake the metering consensus makes. It treats “pay for what you use” as obviously fairer, and therefore obviously better. For the vendor’s margin, it is. For the customer’s nervous system, it is not.

A metered bill is a variable bill, and variable bills create a specific kind of friction: you hesitate before every expensive action. Should I run the agent again, or have I burned enough today? That hesitation is death for a product trying to become a daily habit. The flat fee’s real magic was never the price — it was that you stopped counting. You used the tool freely because using it more cost nothing more. Metering quietly removed that and replaced it with a small tax of anxiety on every click.

So let me be precise about what I think comes back, because it isn’t the naive 2025 deal. I don’t think anyone reopens the all-you-can-eat buffet — truly uncapped use of frontier models at a flat price is exactly the economics that just blew up. What returns is something subtler and more durable: flat-feeling pricing. One fixed monthly bill, headroom generous enough that ordinary heavy use never hits a wall, and the cost controls moved out of the customer’s face and into the vendor’s back end. Not infinity — freedom from calculation. Customers were never really paying for unlimited compute. They were paying to stop doing arithmetic before every click, and that is a product you can sell at a premium.

Which sets up a classic market opening. When every competitor optimises the same variable — here, cost-recovery precision — they all neglect the same thing, and the neglected thing becomes ownable. In a field of meters, the product that says “one price, sensible limits, stop counting” isn’t the cheap option. It’s the premium one, because it sells relief to people who’ve just spent a fortnight watching a number tick up.

The most likely mover: Anthropic, for three reasons

If you’re looking for who fills that gap first, look for who already has the pieces in place. On Anthropic’s side, three line up.

One — it already sells the product, in all but name. Claude’s Max plan is a fixed $200 a month aimed squarely at heavy daily users, and its whole pitch is headroom: one price, big allowance, stop watching the counter. It isn’t literally unlimited — there’s a five-hour session window and a weekly cap, which matters and I’ll come back to — but as a flat-feeling plan for power users, it’s the most developed example in the market.

Two — the economics already work. By Finout’s reckoning, a Max user who genuinely burns through the session limits can consume the equivalent of $600 to $1,500 a month in API tokens for a flat $200. Anthropic can absorb that because the cost controls are quietly built in: rolling-window limits instead of a published token ceiling, cheaper models in the family (Haiku, Sonnet) carrying routine load while the expensive flagship is reserved for what needs it, and an opt-in overage toggle for the rare user who blows past the cap. The fixed sticker price sits out front; the levers that protect margin sit behind it.

Three — it’s already using capacity as a weapon. In May 2026 Anthropic raised Max’s weekly limits by 50% through mid-July, a move widely read as a direct counter to OpenAI’s Codex. A company that responds to competition by handing power users more predictable headroom — rather than metering them harder — is telling you where it thinks the loyalty is. That’s the instinct the prediction rests on.

Why not the others? Microsoft and GitHub just spent the week teaching their users to watch the meter — reversing that narrative now would be whiplash, and their own MAI cost base is aimed at margin, not at a flat-rate pitch. OpenAI has the consumer mindshare but built its $200 tier on the same metered-overage logic as everyone else. Cursor lives closest to the dev workflow but already took its backlash moving the other way, toward credits. Each could move; none is as cleanly pre-positioned as Anthropic, which has the cost base, the meter-averse customer, and a demonstrated reflex to compete on headroom.

What to do about it — before the comeback arrives

This isn’t just a spectator’s prediction. The shift to metered pricing changes how you should behave today, whichever side of the tool you sit on.

If you pay for AI tools: stop optimising for sticker price and start optimising for usage shape. The cheapest plan on paper can be the most expensive in practice if your work is agent-heavy. Price out a real month — your actual tasks, your actual volume — on each option. For consistent daily use, a flat or near-flat plan like Claude Max often beats metering outright; for spiky, occasional use, pure metering wins. The headline number lies; your usage pattern tells the truth.

If you build on AI: predictability is now a feature you can ship. If your product wraps a model, the competitor who exposes raw metered cost to their users is handing you an opening — offer a fixed-price tier with sensible guardrails and you’ll win the customers who’ve been burned by surprise bills elsewhere. The guardrails — model routing, context limits, capped runs — are the same ones the big labs use. You don’t need their scale to copy the structure.

If you’re just trying to keep your own costs sane: treat your context window like cash. Every token you feed a model — the file you paste, the history you carry, the re-run you fire off out of impatience — is now a line item. Trim the input, cap the output, and stop the blind retries. The same discipline that protects your bill under metering is the discipline that makes you faster regardless.

The bet, stated plainly

I’ll put my own view on the table, because a thesis with no falsifiable claim attached is just commentary.

So here is the bet, with terms tight enough to be marked right or wrong. By the end of Q1 2027, at least one of the top five AI coding or assistant tools will launch or relaunch a fixed-price plan — one flat monthly fee, no per-token metering on the core workflow, generous enough that typical heavy use never triggers an overage — and will market it explicitly on predictability rather than capability. Anthropic, via a harder push on Max, is my single most likely candidate, for the three reasons above. What would prove me wrong: the whole field stays on pure usage-based pricing, and the only “flat” plans on offer are thin tiers that meter the moment you actually lean on them. What I am not predicting: a return to the 2025 subsidy, or literally uncapped frontier compute. The product I’m calling is flat-feeling, not infinite — a fixed bill defended by quiet guardrails, sold to people who’ll pay extra never to watch a counter again.

Everyone is racing to meter. The interesting money is on who zags. If you think metering is permanent and predictability never comes back, I genuinely want to hear why — reply to this week’s issue and tell me.