Why AI Pricing Pages Are So Hard to Compare

Put four AI pricing pages side by side and you will find four different units. One sells credits, one sells tokens, one sells seats, one sells generations. None of them converts cleanly into the others, and that is not an accident of an immature market — it is a reasonable response to genuinely unstable underlying costs, with a convenient side effect for the seller.

The four units, and what each hides

Credits

An internal currency whose exchange rate the vendor controls and can change. Credits make a plan look generous while keeping the actual cost per useful action unstated — and different actions usually consume wildly different amounts.

Tokens

The most honest unit and the least usable one. Nobody thinks in tokens. Worse, the number consumed by a given task varies with the input, so a token price does not answer "what will this cost me per document".

Seats

Comprehensible, and increasingly a poor fit. When one person can drive a workload that used to need five, per-seat pricing charges for headcount while the cost is driven by usage. It also quietly penalises the occasional user.

Generations

Clear until you discover that a generation is not a fixed thing. A one-line answer and a twenty-page report can both be one generation, which makes the unit almost meaningless for planning.

Why nobody standardises

Two honest reasons. Underlying costs still move quickly, and a vendor who commits to a stable unit takes on that risk. And products differ enough that a shared unit would flatten real distinctions.

One less honest reason: incomparable pricing is difficult to shop. When no two pages can be lined up, comparison becomes work, and work that a buyer skips is a discount the seller never has to give.

What to do about it

Stop comparing plans and start comparing your own workload. Define one unit that means something to you — a processed document, a published article, a handled ticket. Run that unit through each candidate on the cheapest paid tier. Record what it consumed. Multiply by your real monthly volume.

The result is a number in your currency for a thing you actually do, which is the only comparison that has ever been possible here. It takes an afternoon, and it routinely reorders the shortlist.

What good pricing looks like

Some vendors do this well, and they are worth rewarding: a published cost per representative task, a usage estimator that takes your numbers, hard spend caps rather than surprise overages, and no penalty for the light user. None of that requires a standard unit. It only requires deciding that the buyer should be able to work out the bill in advance.