The AI pricing plateau

There are two ways to ask how AI pricing is changing. The loud way — are prices going up? — gets you a shrug and a few anecdotes. The useful way is structural: when an AI product changes how it charges, which direction does it move, how far, and how often? Ask it that way and the noise resolves into a shape. One force pushes nearly every product the same direction. Most of them get stuck in the same place. And the model everyone assumes is coming has almost no one in it.

I read the public pricing pages of 51 AI products and tracked where each one moved over the last 18 months (between January 2025 and the middle of 2026) — coding tools, image and video generators, support agents, automation platforms, and the incumbents bolting AI onto software you already pay for. Every page archived on the day I read it; the result is the AI Credit Index. This essay is the map that came out of it. The companion piece, how AI hides its prices, is about what that does to you as a buyer. This one is about the physics underneath — why the prices move the way they do.

One continuum, six rungs#

Forget the marketing categories. Every product in the Index sits somewhere on a single line that runs from access-priced to usage-priced — from “a seat buys you everything” to “your usage is the bill.” Six rungs, left to right:

• Rung 1 · Bundled — flat seat or subscription, no meter. You pay, you use, nobody’s counting.
• Rung 2 · Soft meter — there’s an allowance, but running out doesn’t really bite. You’re throttled or nudged, not stopped. This is where “unlimited*” lives.
• Rung 3 · Hard allotment — a real ceiling. Hit it and you’re blocked until the cycle resets or you buy another pack.
• Rung 4 · Metered overage — keep working past the allotment; the meter auto-bills the excess.
• Rung 5a · Per-action — the priced unit is the action itself. Minimal subscription floor; usage is the price.
• Rung 5b · Per-outcome — you pay only for a delivered result: a resolved ticket, a qualified lead.

The line has a direction. The meter only ever grows — from a bolt-on, toward the whole invoice. Nobody walks left. So a product’s rung isn’t just a description of how it charges today; it’s the best single predictor of its next move. That’s the first reason this is worth mapping rather than just listing.

The shape is the finding#

Here’s where the 51 actually sit. The access-priced left has all but emptied: the flat, no-meter models are essentially gone, down from a handful at the start of 2025 to nearly none. Metered overage roughly tripled, from four products to about a dozen. And the usage-priced right — the pure-consumption frontier everyone talks about — holds just three.

Everything else — 34 of the 51 — is bunched on one rung: the hard allotment. A fixed monthly bucket of credits, a real wall when it empties, a pack to buy to keep going. Not a soft “fair-use” meter, not open-ended overage. A ceiling.

That bunching is the whole story. The market hasn’t fanned out across the continuum exploring different models. It has converged — on the same answer, expressed three ways: meter the work, cap the plan, sell the refill. The credit has become the new seat: the unit you buy, ration, and run out of. Software has always played the value-pricing game; the seat was only ever a proxy for the value delivered, and the credit is the newest proxy — a finer one, pegged closer to the work itself.

One engine: compute cost#

Why this direction, and why now? Because for the first time in the history of software, the marginal cost of using the product is large, real, and visible to the vendor. Every model call costs them money. A flat seat price asks the vendor to eat unlimited variable cost on a fixed fee — a bet they lose precisely on their heaviest, most valuable users.

So the meter is, at root, a pass-through. Across the moves I could attribute to a cause, compute cost was the dominant one by a wide margin — the engine under most of the repricings in the panel. This is the honest, sympathetic reading of the whole migration, and it’s the subject of its own essay: AI pricing is just COGS finally showing up. Software is becoming a business with a real cost of goods sold, and pricing is catching up to that fact.

The engine explains the direction. It does not explain why everyone stops in the same place.

The plateau#

Once a product reaches the hard allotment, it tends to stay there — and it’s the strength of that pull, more than the convergence itself, that I didn’t expect. I’d assumed steady rightward migration — soft meters hardening into overage, overage maturing into per-action. Instead I found a plateau that products climb onto and then don’t leave.

The dwell times tell it plainly. The products sitting at a hard allotment have, for the most part, been there for three to four years — Runway since 2021, Synthesia and Surfer and Apollo since 2022. Of the products already on that rung at the start of the window, almost all were still on it eighteen months later. The hard allotment isn’t a way station on the road to consumption pricing. For most products, it’s the destination.

What keeps them there is a brake, and the brake is the buyer. When a product does try to push further right — to harden the meter, to make overage the default — the backlash is immediate and loud, and it works just enough to stall the move without reversing it. The clearest case is the coding tools. When Cursor reworked its $20 plan in mid-2025 so the same money bought roughly half the usage, the revolt produced a public apology and refunds within weeks. When GitHub Copilot re-based its meter to token-level billing, heavy users reported bills jumping many times over, and the complaints followed. But notice what didn’t happen in either case: nobody went back to the flat plan. The model was kept; only its granularity and its manners were softened. Backlash brakes the migration. It almost never reverses it. (The mechanics of those particular moves are their own essay: the price didn’t change, your bill did.)

So you get a sticky equilibrium. Compute cost pushes right; buyer tolerance pushes back; the product settles onto the hard allotment and parks. That’s the plateau.

The sealed frontier#

Now the part everyone gets wrong. The pure-consumption end of the line — paying per action, or better yet per outcome — is treated in most strategy decks as the inevitable future of AI pricing. “Software becomes a service; you’ll pay for results.” The data says the opposite: the frontier is real, it is valuable, and it is almost completely sealed.

Only three products in the Index price purely on consumption — and only two of those charge per outcome. The telling fact isn’t the small number; it’s that every one of them was born there. Intercom’s Fin support agent launched in 2023 charging $0.99 per resolved conversation and has held that model ever since; Zendesk’s resolution-priced agent is the same shape. The third, Salesforce Agentforce, launched at $2 a conversation and has since pulled back to per-action pricing — it reached for the frontier and retreated, which is its own cautionary tale. But notice: not one of the three migrated in from a seat or a credit model, and Agentforce moved only within the frontier, never off it. Inbound traffic from the rest of the continuum: zero.

The reason is measurement, and it’s a hard constraint, not a temporary one. You can only sell an outcome you can both define and prove. “A resolved support ticket” clears that bar — there’s a clean event, both sides can see it, and you can refund when it doesn’t happen. “A better-designed slide,” “a faster codebase,” “a smarter analysis” do not. For the overwhelming majority of AI work, the output is too entangled with the user’s own judgment to bill as a discrete result. So the frontier stays the preserve of the few categories where the outcome is legible — and stays closed to everyone else. This is the spoke for founders deciding where to play: almost no one can charge for outcomes — and that’s the whole game.

Why “punctuated equilibrium”#

Put the three observations together — one engine, one plateau, a sealed frontier — and the pattern has a name. It’s punctuated equilibrium: long stretches of stability, broken by short, sharp jumps, with very little smooth drift in between.

That’s exactly what the Index shows. Products don’t slide gradually rightward. They sit still for years on a rung, then reprice abruptly when compute cost forces the issue — a meter bolted on in a single release, a unit redefined overnight, a tier restructured in one announcement — and then settle into a new equilibrium and sit still again. The continuum is real, but you don’t traverse it at a walk. You jump it, rarely, under pressure, and then stop.

This is why the model is predictive rather than merely descriptive. If pricing drifted smoothly, today’s price would tell you little. Because it jumps and then sticks, a product’s current rung is a strong tell for its next move: the bundled hold-out is one compute-cost shock away from a meter; the soft meter is one enforcement decision from a hard wall; the hard allotment will reach for overage the moment it can survive the backlash; and the frontier will stay the address of the few who were born there. The change log is built to test exactly these predictions, edition over edition.

How to read your own next move#

If you buy these tools: stop pricing the sticker and start pricing the rung. A product on the hard allotment is telling you that the refill, not the subscription, is where your real spend will land — so negotiate the top-up rate now, while you have leverage, not in the busy month when you don’t. A product still on a flat plan is not a bargain; it’s a repricing waiting to happen, and you want a contract that survives it.

If you build them: the plateau is not failure, it’s the equilibrium the economics support — but know that it’s a plateau, not a peak. The honest question isn’t “how do we get to outcome pricing?” (you almost certainly can’t walk there) but “is our outcome legible enough to have been born on the frontier, and if not, how do we run the best metered business on the rung we’re actually on?” Convergence means your pricing is no longer a differentiator. Clarity about it can be.

Pricing has always been where a company decides how much of the value it creates it gets to keep. AI didn’t change that. It just gave the decision a motor — compute cost — and a shape: one engine, one plateau, a frontier almost no one can reach. The whole map, product by product and archived to the date, is the AI Credit Index. It’s free to cite, and it refreshes each quarter — because the next jump is the only part of this that’s hard to see coming.