Busy signals
A proposal for pricing AI subscriptions so that people never watch a meter, heavy use never destroys the margin, and nobody hits a wall in the middle of their work.
The problem
On 1 December 1996, America Online stopped charging by the hour and offered unlimited internet access for $19.95 a month. The company expected usage to rise by about half. It tripled.
What unlimited pricing did to AOL
Average time online per member
AI products have been living through the same story. In June 2025 the coding tool Cursor replaced its fixed monthly allowance of requests with a pool of credits tied to what each request actually cost. Users who ran out sooner than they expected were angry, and within three weeks the company’s chief executive apologised and offered refunds. The next month Anthropic added weekly limits to its Claude subscriptions. It said they would affect fewer than 5% of subscribers.
The underlying problem is easy to state. Every request to a model costs real money. Usage across customers is extremely uneven. A flat price that suits the typical customer loses money on the heaviest ones, and they are often the most committed.
Why the usual answers fail
There are two standard responses, and both have well-documented costs.
Meter everything. Charge for what people use. This is fair and it protects the margin. It also changes how people behave. In 2006 the marketing researchers Anja Lambrecht and Bernd Skiera studied thousands of internet customers and found that many chose a flat rate even when paying by use would have cost them less. One of the causes they identified is the taxi-meter effect: watching the cost tick upward takes away from the enjoyment of the ride. A user who sees a meter starts to ration. They ask fewer questions and try fewer things. For a product that depends on people forming a daily habit, that is expensive in a way that never shows up on the cost line.
Cap it. Keep the flat price and add a limit. The trouble is where the limit bites. It stops your most engaged customers, at the moment they are most engaged, usually in the middle of a task. The product has in effect punished its best customers for using it.
So the requirement looks contradictory. Users should feel they are on a flat rate, the seller should have the economics of a meter, and nobody should hit a wall.
The observation that resolves it
Look closely at what Anthropic said when it introduced its limits. The company did not describe people typing a great deal. It described subscribers running its coding agent “continuously in the background, 24/7.”
That detail points to something general. The extreme usage comes from work that nobody is watching.
A person sitting at a screen is limited. They read at a certain speed, think between requests, eat lunch and go home. However enthusiastic they are, the amount of model usage one attentive person can generate in a day has a ceiling, and that ceiling can be estimated.
Work that runs without a person has no such ceiling. An agent can run all night. Ten agents can run in parallel. A scheduled job can fire every five minutes for a month.
This suggests that usage divides into two kinds, and that they should be priced differently.
Two kinds of usage
| Attended work | Unattended work | |
|---|---|---|
| What it is | A person is at the screen, reading each answer and steering | A task runs in the background, on a schedule, or as one of many in parallel |
| Natural ceiling | Hours in a day and the speed a person can read | None |
| Who is waiting | The user, right now | Nobody, for minutes or hours |
| How to charge | Flat fee, at full speed, with no visible meter | A fixed quote in credits before the task starts |
| When capacity is short | Protect it | Offer to run it later for fewer credits |
The proposal
Part one: attended work is flat, fast, and unmetered
When a person is present and steering, charge a flat fee and show no meter. The flat fee is safe here because human attention sets the upper limit. A team can calculate the cost of a user who works attentively for ten hours a day and set the price above it. Beyond that, the heaviest attended user is simply the best customer.
Part two: unattended work gets a quote
When a user starts a background task, the product states a fixed price in credits before it begins: “This will use about 40 credits. You have 600.” The user accepts or declines. If the task turns out to cost more than expected, the seller absorbs the difference.
Two pieces of research support this design.
The first is about who should carry uncertainty. The cost of a single agent task is hard to predict, even for the seller. One run finds the bug in three steps and another takes sixty. For one user and one task, that uncertainty is large. Across thousands of tasks it nearly disappears.
Uncertainty shrinks as tasks are pooled
How precisely the average cost per task is known. Illustrative, assuming each task's cost varies by as much as its average.
The seller is by far the better party to carry this risk. Token-based pricing does the reverse. It hands the uncertainty to the one person who cannot spread it.
The second piece of research is about how paying feels. In a 1998 paper called The Red and the Black, Drazen Prelec and George Loewenstein showed that people enjoy something more when payment is separated from consumption, and that paying beforehand is less painful than paying during. A quote puts the whole decision at the start, so the user is not watching a meter while the task runs.
Part three: when capacity is tight, offer time
Unattended work has one more property: nobody is waiting for it. That makes it movable. A task that would run at three in the afternoon, when demand peaks, can run at three in the morning.
The model providers already price this. Both Anthropic and OpenAI sell batch processing, for work that can wait up to a day, at half the normal rate. A product can pass that on directly: “Run now for 40 credits, or tonight for 20.”
The cap has now become a choice the user makes. A heavy user who runs out of credits is offered a slower lane or the option to buy more. Electricity companies have long managed peak demand this way, with cheaper rates at night.
How to tell if it is working
- Walls hit. The share of attended sessions that are interrupted by a limit. The target is close to zero.
- Quote accuracy. Actual cost divided by quoted cost, by type of task. It should average near one. If it drifts, the quotes need recalibrating.
- Margin on credits. Unattended work should carry a healthy margin by itself, without being subsidised by the subscription.
- Share moved off-peak. How much unattended work users choose to delay. This is capacity gained without buying hardware.
- Retention of heavy users. The customers that caps used to push away.
Where it could go wrong
People will disguise unattended work as attended. A script can imitate a person. The defence is to look at rhythm. Real attention has pauses and irregular gaps, and a loop does not. The attended tier still needs a fair-use ceiling, set high enough that no real person reaches it.
Quotes will be wrong for new kinds of task. Until there is enough history, quote a range, or meter the first run and quote from then on.
The boundary will blur. A person may start a long task and glance at it every twenty minutes. I would treat anything that runs for more than a few minutes without input as unattended, and say so plainly in the product.
AOL’s problem in 1996 was that every hour online was the same hour, taken from the same phone lines at the same time. AI products are in a better position, because they can tell work that someone is waiting for from work that can wait, and they can price each differently.
Sources
- CNN Money, AOL hikes monthly fee, 1998, and Wired, ‘You’ve Got Payback!’ AOL to Refund User Fees, 1997.
- TechCrunch, Cursor apologizes for unclear pricing changes that upset users, 2025.
- Engadget, Anthropic is rate limiting Claude Code, blaming some users for never turning it off, 2025.
- Lambrecht and Skiera, Paying Too Much and Being Happy About It, Journal of Marketing Research, 2006.
- Prelec and Loewenstein, The Red and the Black: Mental Accounting of Savings and Debt, Marketing Science, 1998.
- Batch processing documentation from Anthropic and OpenAI.