What stops you first is the allowance, not the price

The price is the number you can compare. The allowance is the number that stops you.

Most software pages publish a price for every tier and most of them say nothing about how much you get at each tier, which is why the price is the number buyers compare. Of the 106 companies in our dataset, 14 publish a tiered AI allowance at all, and only 9 of those publish a ladder with more than one number in it.

Where a ladder exists, the tiers are not evenly spaced. The median step between two published tiers is 2.5x, the smallest is 1.6x and the largest is 10.0x. 5 of the 15 steps we could measure are five times or more, and 4 are eight times or more. A buyer who assumes the next tier is twice the work is wrong on a third of the steps we found.

And the top of the range is usually a word rather than a number. 7 of the 14 tiered publishers stop at Unlimited or at custom, which means the part of the ladder where the decision actually gets made is the part you cannot see.

There is a floor problem as well, and it catches people earlier. 1 of the tiered publishers give the lowest tier nothing at all, so the first limit you hit may be the entry plan rather than the top one.

Last verified 2026-10-23. Prices and allowances change without notice. Re-read the table before you rely on a row in it.

The companies that publish a ladder, and what each step is worth

Every vendor in the dataset whose pricing page publishes an AI allowance that changes by tier, with the step between each pair of tiers. The step is the ratio of the published allowances, so it is arithmetic on the vendor's own numbers.

Every published step between two tiers, in one view

The step from one tier to the next, for every ladder in the dataset. A bar chart would be the wrong shape here, because the point is the spread and where the steps cluster.

The median step is 2.5x and the largest is 10.0x. The steps cluster at the bottom and tail off, which means the typical vendor roughly doubles the allowance and the exceptional ones multiply it.

Why the steps are not even, and why that matters

A ladder with even steps can be planned against. A ladder with uneven steps has to be read.

The smallest step we found is 1.6x and the largest is 10.0x, with a median of 2.5x between them. So the typical vendor roughly doubles the allowance when you go up a tier, and the exceptional ones multiply it by five, eight or ten. The exceptional ones are where the overspend happens, because a buyer who has learned to expect a doubling does not budget for a ten times jump.

The step is also not always in the vendor's favour. A ten times jump in allowance for a small jump in price is a generous tier, and a two times jump for a large price rise is a thin one. Without both numbers you cannot tell which you are looking at, and the price is the one you can see.

There is a second problem with an even ladder, and it is the one buyers hit first. The floor. 1 of the tiered publishers give the lowest tier nothing at all, so the first thing you hit may be the entry plan rather than the top one, and the top of the ladder is irrelevant if the bottom of it is zero.

The publishers who give you the floor and hide the ceiling

These vendors publish exactly one allowance and then write unlimited or custom above it. You can see the bottom of the range and not the top.

One vendor shows both ladders, and the gap between them is the product

It is the only row in the dataset where an allowance ladder and a price ladder are published for the same set of tiers.

Streak publishes AI credits per user per month as 20 / 150 / 500, and prices the three tiers at $59 / $69 / $129. The first step up is 7.5x on the allowance and 1.17x on the price, which is 6.41x more AI work per dollar. The second step is the other way round, at 6.41x the work per dollar falls.

That is the whole argument for reading the allowance rather than the price. A buyer who only reads prices sees a cheap middle tier and a normal top tier. A buyer who reads the allowance sees that the middle tier is where the value is and that the top tier is where the value ends.

It is also why we cannot compute a market-wide figure here. One vendor in 106 publishes both ladders, and one case is an anecdote rather than a finding.

Each ladder, with what it publishes

The nine vendors that publish a tier ladder, one at a time, with the unit they meter in and where the top of the range sits. The units are the part that stops buyers comparing them, and each one reads differently.

Nine ladders, nine different units, and no conversion between them

This is the part that stops a buyer from doing the obvious thing, which is to put the nine vendors side by side and find the cheapest.

The nine ladders we found meter in tokens, in credits, in prompts, in questions, in calls, in completions and in actions. A token is a fraction of a word. A prompt is a question with a fixed length. A credit is whatever the vendor says it is, and on the vendors we read, it is usually a number of model calls. An action is a step in a workflow. None of them convert into each other without the vendor's own conversion, and the vendor is the only party who has it.

So the comparison you can make is not which vendor gives you the most for the money. It is which vendor's unit matches the work you are actually doing. A team that runs two hundred short prompts a month and a team that runs twenty long ones will burn the same unit at very different rates, and no price comparison between them means anything.

The second thing the units tell you is where the vendor has put the pricing pressure. Credit-based ladders put the pressure on volume, which scales with your usage and gets worse as you succeed. Per-seat ladders put it on your team, which is predictable and gets worse as you hire. A per-minute or per-incident unit puts it on the hardest thing to forecast. Reading the unit is a faster way to understand a pricing model than reading the tiers.

Four things the ladders say about the category

The findings, in the order they matter to a buyer picking a tier.

What comparing two vendors on this actually looks like

Two vendors, one real workload, and the arithmetic that a pricing page does not do for you.

Vendor A sells AI credits per user per month, in tiers of 20 / 150 / 500, at $59 / $69 / $129. Vendor B meters prompts per tier per month, in tiers of 10 / 100. You have a team, and you expect to run a few thousand prompts a month across it.

Vendor A's ladder, as published, gives you the credits. How many prompts that is depends on the vendor's own conversion, and it is not on the pricing page. So the first thing you have to do is ask, and the answer will tell you which tier you land on.

Vendor B's ladder tells you the tier directly, because it is counting the same thing you are counting. Which means Vendor B is the easier comparison and Vendor A is the cheaper one, and the gap between those two statements is the whole buying decision.

This is why the unit is the first thing to read and the last thing to assume. A comparison page that puts two vendors in a table without telling you the unit is not a comparison, it is a list. A comparison that tells you the unit, and shows you where each one's ladder stops being a number, is a decision you can make.

Four questions to ask before you pick a tier

The gap is a disclosure gap, and these are the disclosures that close it.

The five questions this study keeps getting asked

Four readers, and what each of them should take

The same measurement reads differently depending on what you are doing with it.

You are buying. You know which number will stop you, and now you know that the number on the pricing page is usually not it. Ask for the allowance ladder and price the step, not the tier.

You are selling. Fourteen of a hundred and six companies publish a tiered AI allowance, and seven of those stop at a word. Publishing the allowance ladder would put you ahead of most of the market on the one question buyers are asking and cannot answer.

You are writing the comparison. A table of prices without the units is not a comparison, and the units are the part that decides the answer.

You are modelling unit economics. The allowance is the variable that scales with your success, and it is the one your revenue model will not see coming if you price it off the tier name rather than the published number.

What this does not tell you

Where this census sits in the rest of our work

Three studies that bear on the same question: what a vendor's own page lets you find out before you talk to them.

How we checked

The evidence is the 106-company hand-coded AI pricing dataset, collected in July 2026 and published at /research/b2b-saas-ai-pricing. Each record holds a billing-model code and a verbatim note taken from that company's own pricing page, so every figure here traces to a sentence a vendor wrote.

We re-coded the dataset for the shape of the allowance rather than the shape of the price. A ladder is a published per-tier allowance: a number of prompts, credits, tokens, questions, calls or actions that changes by tier. A single published allowance with no tiering is a single, a price ladder with no allowance is a price ladder, and an allowance expressed only as a rate is a unit price. The step between two tiers is the ratio of the published allowances, and every step factor on the page is computed from the numbers rather than read off the page.

The curated coding file carries the verbatim note for every row, and the analysis refuses to run if any of them does not match the record it came from, so the coding can be checked against the source. Two boundaries are worth naming. A vendor that publishes one number and then writes unlimited or custom above it is counted as a floor with a hidden ceiling rather than as a ladder, because the step between two numbers cannot be computed from one. And the work-per-dollar figure comes from a single vendor, because it is the only one in the dataset publishing both ladders.