AI usage costs: where does it add up?
Most AI tasks cost pennies. Some cost a lot more, and they are not always the ones you would expect. Here is how AI usage is charged, what typical business tasks cost to run, and what turns a small bill into a large one.
When businesses budget for AI, they usually budget for building it. The cost that surprises them comes later: the monthly bill for using it. A pilot that costs a few pounds a week in testing can cost hundreds a month once a whole team relies on it, or once it starts handling tasks automatically.
How AI usage is charged
There are two ways to pay. The first is a subscription: a flat monthly fee per person for a tool such as ChatGPT, Claude or Copilot. That is simple and predictable, and right for general staff use.
The second is paying for use. When AI is built into your own systems, answering from your documents, processing invoices or running behind a website, you pay the model provider for each piece of work it does. This is where costs vary, and where to watch out.
Usage is measured in tokens. A token is a fragment of text, roughly three-quarters of a word. You pay for the tokens you send in (your question, plus any documents or instructions attached to it) and for the tokens the model sends back. Three details matter more than the headline price:
- Output costs more than input. On current price lists, the words a model writes typically cost around five times as much as the words it reads.
- Hidden thinking is billed as output. Many newer models “reason” before they answer. You never see that working, but you pay for it at the output rate.
- The whole conversation is resent each time. On an AI chat, models don’t remember. Every follow-up question sends the earlier exchange again, so long conversations get steadily more expensive per message.
Bar lengths show relative cost. Based on September 2026 list prices; illustrative only.
Three tiers of model
Every major provider now sells models in roughly three tiers. Large frontier models are the most capable and the most expensive. Mid-tier models handle most business tasks well at a fraction of that. Small models are very cheap and suit simple, high-volume work.
The gap is wide. On published price lists in September 2026, a typical frontier model cost around two and a half times a mid-tier model, and roughly twenty times a small one, for the same amount of text. Choosing the right tier for each task is the single biggest lever on cost.
What typical tasks cost to run
The table below estimates monthly running costs for six tasks a 50 to 100-person business might automate. Each figure is shown for a small, mid-tier and frontier model, so you can see how much the choice of model changes the answer.
| Task and assumed volume | Small | Mid-tier | Frontier |
|---|---|---|---|
| Summarising tenders or contracts20 long documents a month | under £1 | £3 | £7 |
| Sorting and summarising incoming email3,000 emails a month | £1 | £11 | £28 |
| Staff questions answered from your documents20 staff, 10 questions a day | £5 | £50 | £126 |
| The same, with heavy reasoning switched onsame volume | £20 | £176 | £441 |
| Website enquiry assistant2,000 conversations a month | £13 | £126 | £315 |
| Multi-step invoice matching1,000 invoices, about 25 steps each | £41 | £394 | £984 |
Illustrative only. Based on published list prices for representative models in each tier in September 2026, converted approximately to sterling, with no caching or bulk discounts. Your figures will depend on document length, how much the model writes, and how prices move.
Two things stand out. Summarising long documents is cheap even on the best models, because the volume is low and the model mostly reads. And the most expensive row is not the one with the most text to read. It is the one where the AI takes many steps for every item.
Same tier of model, very different bills. The two most expensive tasks are the ones that make the model write or think a lot, not the ones with the most to read.
What is cheap
- Reading, even lots of it. Summarising, classifying and extracting details from documents is mostly input, which is the cheapest part of the bill.
- Simple, repetitive jobs on small models. Sorting emails, tagging documents and pulling fields from forms rarely need a large model.
- Work that can wait. Most providers offer around half price for jobs processed in bulk rather than instantly, such as an overnight run.
- Repeated instructions. If every request starts with the same long instructions or reference material, providers can store it and charge a fraction of the normal rate to reuse it.
- Low-volume, high-value tasks. Twenty tender summaries a month on the best available model still costs less than a sandwich.
What is not
- Multi-step automation. A system that looks things up, decides, acts and checks may call the model dozens of times per item, each time resending what it has gathered so far. Cost scales with steps, not with items.
- Heavy reasoning on routine questions. Thinking modes improve hard answers but can multiply the cost of easy ones several times over, as the table shows.
- Frontier models used by default. Running everything on the most capable model is the most common avoidable cost. It typically adds two to three times the bill compared with mid-tier, for little visible difference on everyday tasks.
- Long conversations. A customer chat that runs to twenty messages resends the whole history each time. The last messages cost far more than the first.
- Open-ended public use. A tool on your website is used as much as visitors choose to use it. Without limits, you do not control the volume.
- Retries and failures. If outputs regularly fail a check and are re-run, you pay twice.
The £2,400 a month question tool
A business builds a tool so staff can ask about products and procedures. To make sure the answer is always available, every question is sent with the full reference pack: around 110,000 words of manuals and specifications. The demo answers are excellent and testing costs a few pounds. Live, with 20 staff asking 10 questions a day on a frontier model, the bill reaches roughly £2,400 a month. Nothing is broken; the model is simply reading the whole library to answer each question.
It could have been avoided by:
- Preparing the knowledge base properly, so documents are split into sections, out-of-date versions removed and everything indexed. Each question then sends only the few passages that answer it, and answers usually improve.
- Using a mid-tier model, which handles this kind of question well.
- Estimating cost at live volumes before launch, not from test usage.
Together, those bring the same workload down to tens of pounds a month.
Keeping the bill under control
None of this is a reason to avoid AI. It is a reason to estimate running cost before building anything, the same way you would estimate the running cost of a vehicle or a machine. In practice that means four habits.
Match the model to the task, and be willing to use different models for different steps of the same job. Estimate monthly volume and cost before you start, using real examples of your documents and questions. Set spending caps with the provider and limits on public-facing tools. And review actual usage after the first month, because real use rarely matches the plan.
Prices also move, mostly downwards. Models that cost a certain amount this year are often matched by cheaper ones within months. A system built so that the model can be changed easily lets you take advantage of that without rebuilding anything.
About the figures: estimates are calculated from providers’ published per-token list prices for representative small, mid-tier and frontier models in September 2026, using stated assumptions about document length and output. Prices change frequently, so treat the ratios between tiers and tasks as more durable than the pound amounts.
More insights
Book a free opportunity review.
A focused conversation about how your business runs and where AI could help — and where it would not.
