Why did our bill jump without our usage changing?

Check the four usual causes: conversations getting longer, attachments getting larger, a switch to a more capable model, or a feature that now works through a model rather than a rule. A change to any one of those moves the total without anybody feeling they are using the tool more.

Which is dearer, what we send or what comes back?

Producing text is generally charged at a higher rate than reading it, and reading is usually the larger quantity in document-heavy work, so both matter and for different reasons. The practical reading is that long attachments and verbose answers are two separate habits worth curbing separately.

How do we estimate before committing?

Run fifty genuinely representative cases, note what they consumed, and multiply by the volume you actually expect rather than the volume you are testing at. That arithmetic is unglamorous and it is the single most reliable defence against a pilot that looked free turning into a line somebody questions.

Glossary

Inference cost

Inference cost is what running a model to answer one request consumes in hardware time, which is what a plan is metering when it caps your messages or rations the strongest option.

In plain terms

Every answer consumes real hardware for a real number of seconds, and somebody pays for that. It is the reason a plan counts your messages, the reason the best model has a smaller allowance than the quick one, and the reason a proposal to run something over a million records is a budget conversation. Nothing about this is a pricing trick; it is what the underlying activity costs.

Why it matters

It explains almost every restriction you meet and turns them from arbitrary annoyances into predictable ones. Once you know the cost of an answer rises with how much was read and how much was written, the pattern of caps, model tiers and slowdowns across every product you use stops looking like a series of unrelated decisions. It also tells you where to look when a bill surprises you, which is nearly always the same four places.

How it works

Cost tracks the amount of text going in and coming out, and the two are usually priced differently. That is why an identical question costs more at the end of a long conversation than at the start: the whole thread is read again each time, so the reading grows while the answer stays the same length.

Model size multiplies it. A larger model consumes more hardware for the same work, which is the reason plans give you a generous allowance of a quick model and a small one of the capable model rather than a single number covering both.

Deliberation is charged even where it is hidden. A model that works through a problem before replying generates a great deal of text you never see, all of it billed, which is why a considered answer can cost several times a direct one with a visible reply of the same length.

The four levers are always the same. Send less, ask for less back, use a smaller model where it clears your bar, and reuse repeated material where the provider supports caching it. Anything more sophisticated than those four is usually a refinement of one of them.

The same question, asked twice

Two identical questions, two very different costs, and nothing about the user's behaviour distinguishes them. In a fresh conversation the model reads a sentence and writes a paragraph. Twenty turns later it reads everything said so far before it can write the same paragraph, because that is how a thread is followed at all. The consequence is the single most useful cost habit available to anybody: starting a new conversation when the subject changes is free, takes a second, and removes an accumulating charge nobody was watching. The same arithmetic explains attachments. A document sent with every message in a thread is read with every message in that thread.

Seen in the wild

Send one request through a routing interface that reports the cost beside the answer, then send the same request with a long document attached and compare.
OpenRouter
Ask an identical question in a fresh conversation and again twenty turns into a long one, and notice which of the two the length of the thread affected.
ChatGPT
Put a model step into an automation running over every incoming record, which is where a cost per answer becomes a cost per month.
n8n

Common misconceptions

People assume

The limits on our plan are arbitrary.

In fact

They are usually a fairly direct expression of what the underlying activity costs. Message caps, tiered model allowances and slowdowns after heavy use all trace back to the same fact, which is why they appear in similar shapes across products from unrelated companies.

People assume

Costs will fall enough that this stops mattering.

In fact

Prices per unit have fallen considerably and use has grown faster, so bills have generally gone up rather than down. Falling unit costs mostly change what becomes worth doing rather than removing the need to think about volume, which is a better outcome and a different one.

Questions

Why did our bill jump without our usage changing?: Check the four usual causes: conversations getting longer, attachments getting larger, a switch to a more capable model, or a feature that now works through a model rather than a rule. A change to any one of those moves the total without anybody feeling they are using the tool more.
Which is dearer, what we send or what comes back?: Producing text is generally charged at a higher rate than reading it, and reading is usually the larger quantity in document-heavy work, so both matter and for different reasons. The practical reading is that long attachments and verbose answers are two separate habits worth curbing separately.
How do we estimate before committing?: Run fifty genuinely representative cases, note what they consumed, and multiply by the volume you actually expect rather than the volume you are testing at. That arithmetic is unglamorous and it is the single most reliable defence against a pilot that looked free turning into a line somebody questions.

Key takeaways

Every answer costs real hardware time, which is what nearly every plan restriction traces back to.
Cost tracks text in and text out, so long threads and large attachments raise it without extra questions.
Larger models cost more per answer, which is why allowances are tiered rather than pooled.
Hidden deliberation is billed exactly like visible output.
The four levers are send less, ask for less, use a smaller model, and reuse what repeats.

Tools that use this

OpenRouter
Cost reported beside the answer, so the effect of an attachment is visible.
ChatGPT
The same question early and late in a thread, where the reading grew.
n8n
Per-answer cost becoming per-month cost, on every incoming record.

Last checked July 2026

All glossary terms