How do we budget for it?

Measure fifty representative cases, note what they consumed, multiply by the volume you genuinely expect, and present the answer as a range. Then set a cap somewhat above the top of that range, so the forecast being wrong becomes a conversation rather than an incident.

Can we cap what it spends?

Usually yes, and it is the most valuable few minutes available on the whole subject. Providers commonly offer both a hard limit and an alert at a chosen threshold. Set both, set the alert well below the limit, and make sure it reaches somebody who would act on it rather than an unread mailbox.

When is a subscription the better arrangement?

When a person is using a tool steadily through the working day, which is the pattern subscriptions are priced for. The meter earns its place when work is intermittent, seasonal, or done unattended at volume. Ending up with both, for different jobs, is a sensible outcome rather than a muddle.

Glossary

Pay-as-you-go

Pay-as-you-go means being charged for what you actually use rather than a flat fee, which suits demand that comes in bursts and makes next month's figure genuinely harder to predict.

In plain terms

A meter rather than a season ticket. Quiet weeks cost almost nothing and busy weeks cost more, which is fair and is also the difficulty: nobody can tell you in advance what a busy week will look like. The two questions worth answering before adopting it are what a bad month would cost and what stops it going further than that.

Why it matters

It is the arrangement behind most serious use, because a subscription per person stops making sense the moment work is being done unattended rather than by a person at a keyboard. It also changes who needs to understand the cost. Under a subscription, finance approves a number once. Under a meter, the decisions that move the bill are made by whoever writes the automation, which is a different person with different information.

How it works

You are charged for units of activity, and the units are usually amounts of text read and produced rather than requests. That is why a small number of large jobs can cost more than a large number of small ones, and why the intuitive measure of how much you used the tool is a poor predictor of the figure.

Nothing is charged while nothing is happening, which is the genuine advantage. Seasonal work, occasional batches and experiments that might come to nothing all cost what they actually cost, with no commitment sitting there in months when nobody touches it.

Forecasting is the real difficulty and it does not go away with experience. The honest method is measuring a representative sample and multiplying by expected volume, then treating the result as a range rather than a figure, because the things that move it, thread length and attachment size, are the things people change without noticing.

Spending caps and alerts are the control that matters and the one most easily left unset. A cap turns a runaway automation into an interruption instead of an invoice, and an alert at a sensible fraction of your expected month turns a surprise into a Tuesday conversation. Both take a few minutes and are usually available.

Where your pattern of work sits, and which arrangement it wants

Neither arrangement is the better deal in general, because they are priced for opposite ends of one line. Toward the left a meter costs nothing in the months nothing happens, which is exactly right for seasonal work and for experiments that may come to nothing. Toward the right a subscription is cheaper and has the additional virtue of a number finance can approve once and forget. The reason to know where you sit is that work migrates along this line: a tool bought for a few people at keyboards quietly becomes an automation processing every incoming record, and the arrangement that was right at one end is rarely right at the other. Most organisations should end up with both, for different jobs. The mistake is not choosing wrongly but choosing once.

Seen in the wild

Send a handful of requests through an interface reporting cost per request, which is the arrangement at a scale small enough to feel.
OpenRouter
Build an automation that runs on every incoming record and work out what a busy week of those records would cost before switching it on.
n8n
Compare a month of one person's ordinary assistant use against the same work run unattended, which is where the two charging models diverge.
ChatGPT

Common misconceptions

People assume

It is cheaper than a subscription.

In fact

It is cheaper for uneven or light use and dearer for steady heavy use, which is what a meter does. For a person at a keyboard most of the day a subscription is generally better value. For work that happens in bursts or runs unattended, the meter usually wins comfortably.

People assume

A mistake could not cost very much.

In fact

An automation in a loop, a job pointed at the wrong folder, or a leaked key will all keep spending until somebody notices, and nobody watches an invoice hourly. This is exactly what a spending cap exists for, and it takes about two minutes to set on most accounts.

Telling them apart

Pay-as-you-go vs Free tier

Pay-as-you-go

A meter with no ceiling. You can always do more, and doing more always costs more.

Free tier

A ceiling with no meter. You can do a fixed amount for nothing and then something changes.

One constrains your budget, the other constrains your work. A trial on a free tier tells you almost nothing about what the metered arrangement will cost you at real volume.

Questions

How do we budget for it?: Measure fifty representative cases, note what they consumed, multiply by the volume you genuinely expect, and present the answer as a range. Then set a cap somewhat above the top of that range, so the forecast being wrong becomes a conversation rather than an incident.
Can we cap what it spends?: Usually yes, and it is the most valuable few minutes available on the whole subject. Providers commonly offer both a hard limit and an alert at a chosen threshold. Set both, set the alert well below the limit, and make sure it reaches somebody who would act on it rather than an unread mailbox.
When is a subscription the better arrangement?: When a person is using a tool steadily through the working day, which is the pattern subscriptions are priced for. The meter earns its place when work is intermittent, seasonal, or done unattended at volume. Ending up with both, for different jobs, is a sensible outcome rather than a muddle.

Key takeaways

You are charged for activity, so quiet periods cost almost nothing and busy ones cost more.
Units are usually amounts of text, so a few large jobs can beat many small ones.
Forecasting is genuinely hard, and the honest output is a range rather than a figure.
A spending cap and an alert take minutes to set and are the easiest controls to leave unset.
It suits burst and unattended work; a subscription suits a person at a keyboard all day.

Tools that use this

OpenRouter
Cost per request, which is the meter at a scale you can feel.
n8n
Unattended work, where the meter runs without anybody watching.
ChatGPT
The subscription pattern, for comparison against the same work automated.

Last checked July 2026

All glossary terms