Credits are how AI hides its prices

For twenty years, buying software meant buying a seat. You knew the number: so many dollars, per user, per month. You could put it in a spreadsheet, compare two vendors, and forecast next year. AI is quietly ending that. The new unit isn't the seat — it's the credit, the token, the action, the "outcome" — a meter that ticks down as you work. And the moment pricing became a meter, the price stopped being a number you could read.

I spent time reading and analyzing the public pricing pages of 52 AI products — coding tools, image and video generators, support agents, the incumbents bolting AI onto software you already pay for — and recording exactly what they say, archived page by page. I built it into something I’m calling the AI Credit Index. The short version: credits are how AI products hide their prices, and the hiding is getting better.

A credit is not a unit#

Start with the number everyone wants — what does a credit cost? Across the panel, the median works out to about a cent and a half. That figure is worse than useless, and it’s worth understanding why.

Because the same word, “credit,” buys wildly different things. Published bundled rates in the Index run from less than a thousandth of a cent per credit to twenty-five dollars per credit — a spread of more than thirteen million to one. In one product a credit is a single generated sentence; in another it’s a finished video clip; in a third it’s an entire resolved support ticket. A “credit” is a currency each vendor mints privately and never has to peg to anything. Comparing the price of one company’s credit to another’s is like comparing the price of one country’s banknote to another’s without the exchange rate.

This is the first thing the meter does: it makes cross-shopping arithmetically impossible. You cannot line up two AI tools the way you could line up two seats. The unit was designed not to travel.

The meter you can’t read#

If you can’t compare credits across products, the next best thing is to compare a product to itself — what does it charge to refill the meter versus what the plan implies? That’s a fair, unit-free question. The problem is you usually can’t answer it either, because the refill price isn’t published.

Of the products in the Index that sell top-ups, only about three in five tell you what a top-up costs anywhere on their public site. For the rest, the price appears only at the in-app purchase screen — after you’ve subscribed, inside the product, where no procurement team comparing options and no crawler indexing the web can see it. The economics of running over are the single thing vendors are least likely to print.

And that’s just the companies that publish a pricing page at all. A whole tier of category-leading AI products — thirteen of them on the Index’s watchlist, the ones selling AI to lawyers, doctors, and enterprises — publish no price whatsoever. “Contact us” is the entire pricing strategy. The refusal to post a number is itself the finding.

Topping up is where they get you#

When you can see both numbers, the picture splits in a way that tells you what you are buying. About half the comparable plans charge more per credit to top up than the bundle implies — and the steepest premiums cluster in the pure-play generators, the products where the credit is the business: Midjourney charges roughly double to refill, and Figma’s cheapest AI tier about four-and-a-half times. The other half price refills at or below the bundle, and these are overwhelmingly the incumbents and suites that bolted a credit onto something you already pay for — Adobe Firefly discounts top-ups about 20% on most tiers, monday and Gamma’s entry plans the same. When the meter is the whole business, the refill is the profit center; when it is an add-on, the vendor has little reason to gouge and sometimes a reason to discount, to keep heavy users from leaving. The median across everyone lands near parity — but the median is the least interesting number here. What it hides is the rule: the steepest premiums fall on the cheapest entry tiers, where the buyer has the least room to negotiate.

Sit with the asymmetry, because it’s the whole game. You choose a plan based on an average month. You buy top-ups in your busiest month — the launch week, the board deck, the customer escalation — which is exactly when you’re least price-sensitive and most locked in. The bundle is the advertised price; the top-up is the real one, and it’s quoted to you at the worst possible moment.

The pricing experts I’ve had on the show keep returning to the same idea: a price is a signal about value and fairness, not just a number (Jean-Manuel Izaret made this the center of our conversation). A top-up premium charged only when you’re cornered is a deliberate choice about when to extract value. It’s good tactics. It’s worth seeing clearly.

The cent that hides a moving number#

Here’s the move that should worry a CFO most, and it hides in plain sight.

A cluster of incumbents — GitHub Copilot, Atlassian’s Rovo, monday.com, HubSpot — have each independently settled on the same convention: one credit equals exactly one US cent. A round number, easy to reason about. Reassuring, even. But pegging the dollar value of a credit says nothing about how many credits an action costs. And that second number — the credit burn rate, how many credits an action consumes, not a company’s cash burn — is the one the vendor can change unilaterally, with nothing on the pricing page moving.

This isn’t hypothetical. As GitHub Copilot shifts to consumption-based billing, users are reporting that the same work now burns far more than it used to, at an unchanged headline rate. The price went up; the price tag didn’t. No competitor’s price-tracker, no procurement benchmark, no renewal review built around the sticker price will ever catch it, because they’re all watching the dollar figure — and the dollar figure is exactly what was held still. It’s the most elegant form of a price increase ever invented: inflation with a fixed label — digital shrinkflation. (I’m logging these as “credit burn-rate repricing” events in the Index, because someone should.)

What happens when you run out is usually unsaid#

One more question any buyer should ask: when the meter hits zero, what happens? Do I stop? Slow down? Get billed automatically?

Across roughly two hundred plan tiers I read, the most common answer was no answer at all — the pricing page simply doesn’t say. Where it is stated, a hard stop is most common, but a meaningful share default you into automatic overage: keep working, get billed, find out at month’s end. Included credits almost always expire each cycle; they rarely roll over. So you’re encouraged to buy a comfortable monthly allotment, you forfeit whatever you don’t use, and if you go over, the tool may keep charging without asking. Every default points the same direction.

How to buy around it#

None of this means credits are a scam. Usage-based pricing is often genuinely fairer than seats — you pay for what you use, and a meter passes through real, variable model costs. The problem isn’t the meter; it’s the opacity around it. So price the opacity out — and how hard you can push depends on how you buy. Self-serve, you can’t negotiate; your leverage is choosing well, sizing right, and watching the meter. Buying for a team, the same moves become contract terms you can actually demand:

• Find the top-up rate before you commit. Self-serve: if it only appears in-app, surface it during the trial and weight that opacity in your choice. Enterprise: make the refill rate a contract term — silence until you’re a customer is the answer.
• Translate credits into your work, not theirs. The marketing example (“1 credit = 1 image!”) is chosen to flatter. Instead, during a trial run ten of your real jobs — the actual support replies, the actual edits — and watch the meter fall. That’s your true credits-per-task. Multiply by your real monthly volume and compare to the plan’s allotment: if your team’s 300 tickets a month burn three plans’ worth of credits, you’ve learned you’re a top-up customer — and at what price — before signing, not after.
• Size to a spiky month. Forecast your worst week, not your average one — that’s when the expensive top-up math actually bites.
• Get exhaustion behavior in writing. Hard stop, throttle, or auto-overage? “Auto-overage, billed by default” is a budget you don’t control.
• Re-check the credit burn rate every quarter. A static dollar price is not a static price. Track credits-per-action over time; that’s where the quiet increases live.

I built the AI Credit Index so this isn’t a matter of vibes. Every figure in it comes from a vendor’s own public page, with the derived math — effective rates, top-up premiums, the disclosure rate itself — computed rather than asserted. It’s free to cite, and it refreshes each quarter, because the prices that matter most here are precisely the ones designed not to sit still.

Pricing has always been where companies decide how much of the value they create they get to keep. AI didn’t change that. It just moved the decision somewhere you have to work harder to see. Worth the look.