Guide
When Does a Flat AI Plan Start Losing Money on One Customer?
An outside-in estimate from 20 public AI software plans. Every break-even figure here is an estimate, not a real bill.
An underwater customer is one who costs more to serve in AI usage than they pay you. Their subscription looks fine on the revenue report.
We wanted to know how much usage it takes to get there. So we read the pricing pages of 20 AI software products and worked out the break-even for one plan from each. We aren't naming them, because these are our guesses about costs we can't see.
The finding
What puts a flat price at risk is a cost that grows with time, such as meeting hours. Using AI isn't enough on its own.
On our central assumptions, all four flat plans selling unlimited AI meeting notes went underwater at 3 hours of meetings a working day. Their break-even was 21 to 42 hours of meetings a month. All three flat plans built around short text or drafts stayed above water in every case we ran, even with no numeric limit. The lowest break-even of the three was about 1,300 outputs a month.
So the word "unlimited" can carry very different risk. Our estimated cost for one unit of usage ran from about $0.001 for a short rewrite to about $9 for a minute of fully generated video, a gap of nearly 8,000 times.
Of the 19 monthly-priced plans, 9 went underwater at heavy use on central assumptions. Cheaper models brought that down to 3 and stronger models took it up to 11.
What we saw, by plan shape
| Shape of plan | Number of plans | Break-even (estimate) | Plans underwater at heavy use: cheap, central, dear case |
|---|---|---|---|
| Flat seat with unlimited AI, sold on meeting hours | 4 | 21 to 42 meeting hours a month | 3, 4, 4 |
| Flat plan with unlimited or uncapped AI, for short text or drafts | 3 | About 1,300 outputs a month or more | 0, 0, 0 |
| Flat seat with no public limit, for long-context AI | 2 | About 290 to 850 long-context tasks a month | 0, 1, 2 |
| Stated allowance of credits, hours or conversations | 10 | Varies, see below | 0, 4, 5 |
| Pay per outcome | 1 | About 9 conversations for each billed outcome | Not counted |
Three things stand out.
Cost that runs by the minute matters more than the length of the answer. A meeting summary may only be a page, but the model has to read about 12,000 tokens to write it, and speech to text charges for every minute. In our sums speech made up most of the cost of a meeting hour. A short rewrite has neither problem.
A stated allowance mostly works. In 6 of the 10 allowance plans, someone using their full allowance stayed below break-even on central assumptions. The lowest of them came to about 2% of what the plan costs.
The other four are worth a closer look. One sells video by the minute. If every second of its allowance were fully generated at a mid-range list rate, the customer would cross break-even. In practice much of a finished video won't be fully generated, so treat that as the ceiling.
The next two were shown to us at rupee prices, while model rates are set in dollars. That can move a plan from safe to exposed with the allowance untouched. One is a confirmed regional plan. For the other we only saw the rupee price, and we couldn't confirm a dollar price or whether it's regional. On central assumptions the full allowance costs about $6 on both, so any price above that clears it. Both estimates are low confidence, because these plans do more than the one thing we priced.
The fourth sells credits for agent work. It only crosses break-even if a credit uses about as much as we guessed, which is 50,000 tokens in and 3,000 out. That's the weakest of the four.
Pay per outcome moves the risk around. You only earn when an outcome is billed, but you pay for every conversation. On central assumptions, a customer needs an outcome rate above about 11.5% to cover the cost of those conversations. The rates vendors publish on their own pricing pages were much higher, at about 50% to 79% on the pages we read. If those hold, the exposed accounts are the ones with low resolution rates, and the typical account is fine.
Method
Prices. We read each live pricing page on November 4, 2026, took one self-serve plan per product, and reopened every page on November 5. Where annual billing was shown, we used the annual price. We converted rupees to dollars at the European Central Bank reference rates for November 3.
Usage unit. We picked the unit that drives the cost for that product: an hour of meeting, a generated minute of video, an AI reply, a credit.
Cost per unit. For text, we assumed tokens in and out for each unit and multiplied by public model rates from Spendline's model price list. 100 tokens is about 60 to 80 English words. Short tasks used a small model in the central case. Long-context and agent tasks used a mid-tier model. The cheap and dear cases moved one tier down and one tier up.
Here's a worked example. An hour of meeting at 150 words a minute comes to about 12,000 tokens in, and we assumed 800 tokens out. On Claude Haiku 4.5 at $1 in and $5 out per million, the summary costs $0.016. Live speech to text at Deepgram's list rate of $0.0077 a minute adds $0.462. That's about $0.48 an hour in total. A made-up $12 seat would break even at about 25 hours a month.
For video we used Runway's API list rates of $0.08, $0.15 and $0.30 a second. For speech we used public text to speech API prices of $0.01 to $0.08 per 1,000 characters, at about 1,000 characters a minute.
Break-even. Monthly plan price divided by cost per unit. This counts AI cost only, the same narrow view as customer revenue minus attributed AI cost. If other delivery costs take 20% of revenue, every break-even comes down by the same 20%. We took that figure as a starting assumption from our margin calculator. It isn't a benchmark.
Heavy use. For plans with an allowance, heavy means using all of it. For the rest we chose a level one busy person could reach, such as 3 hours of meetings or 20 long-context tasks a working day. Change these and the counts change.
Honest limits
- No company publishes its credit-to-token rate. Some publish credits per reply or per minute. None publish what a credit costs them. Every token count here is our assumption.
- We used list rates. A company at scale may pay less, and one running its own speech model could pay far less. If speech cost a fifth of list, meeting plan break-evens would move from 21 to 42 hours to about 92 to 184 hours, and our heavy user would no longer be underwater.
- Prompt caching isn't included. It can cut the cost of long repeated inputs a lot, which would help the long-context plans most.
- Some costs are missing too. Storage, search and rendering aren't counted.
- Plans are bundles. The seat price also covers everything that isn't AI, and a quiet seat offsets a busy one. One underwater customer says nothing about the health of the whole business, and we make no claim that any company here loses money overall.
- Limits can be unpublished. A plan with no number on its pricing page may still have a fair-use policy.
- Twenty plans is a small sample, chosen by hand. It shows patterns and can't give rates for the market. We didn't verify any company's real costs, and nine of the 20 estimates are low confidence.
- We have no customer data behind this. Spendline has no paying customers yet, so none of these figures comes from a real bill.
What to do with it
Run this for your own plan. Divide your price by your cost for one unit of the thing that drives your bill, then ask how many of your customers are past that number. Seeing the total provider bill is easy. Working out which customer it belongs to is harder.
The margin calculator gives a rough split between your heaviest customers and everyone else. Spendline does this on routed traffic: it records cost per call by customer, compares it with revenue where you share revenue data, and checks budgets before a call goes out. If you'd like to see it on your own usage, we run a guided 30-day free pilot.
FAQ
What is an underwater customer?
A customer who costs more to serve in AI usage than they pay you. Their subscription looks fine on the revenue report, so you need cost per customer to spot them. It's a statement about one customer and doesn't mean the company is unprofitable.
How do I work out the break-even for my own plan?
Divide your monthly price by your cost for one unit of the thing that drives your bill. That gives the usage at which a customer has used up their whole subscription in AI cost. Then count how many customers are past that number.
Does a usage allowance stop customers going underwater?
Mostly. In 6 of the 10 allowance plans we looked at, someone who used it all stayed below break-even on our central assumptions. The exceptions came from a costly unit, a local price set against dollar model rates, or a heavy assumption about what a credit uses.
Are these real cost figures?
No. They're our estimates from public prices. No company publishes its credit-to-token rate, we didn't verify anyone's real costs, and Spendline has no paying customers, so none of this comes from a real bill.
Sources
Written by Aary Patnaik, Spendline co-founder. Spendline sells a product that tracks AI cost per customer, so we have an interest in this question. Sources are linked where they're used, and vendor pricing pages were read on November 4 and 5, 2026. If you think we've got a sum wrong, email fida@spendline.ai and we'll correct it.
See what this would look like in your product.
Start with a guided 30-day free pilot on live production traffic. The 30 days start when your traffic flows through Spendline, not at sign-up. We agree the scope, success criteria and paid terms before you start. There is no Spendline subscription fee during the pilot; your model-provider charges still apply.
Prefer email? fida@spendline.ai