AI models: what the differences are, what they cost, and how to choose one
A plain-English guide for managers early in their AI journey. No jargon required, and no technical background needed to make a good decision.
When people in business say "AI", they usually mean one product they have tried, most often a chatbot. Behind that product sits a model: the engine that reads your question and produces an answer. There are dozens of these models, made by a handful of companies, and they differ in capability, speed, price and where they run.
Understanding those differences is one of the most useful things a manager can learn early on. The choice of model is where most of the running cost is decided, and a good deal of the risk. It is also a choice you can make well without being technical.
Who makes the models
A small number of providers build the largest models: OpenAI, Anthropic, Google and a few others. These are reached over the internet, either through the provider's own app or through a connection (an API) that lets other software use them.
Alongside them sit open-weight models, released by companies such as Meta and Mistral and by a growing number of research labs. These can be downloaded and run on hardware you control, including inside your own cloud tenancy.
Then there is a layer of cloud platforms, chiefly Microsoft Azure, Amazon Web Services and Google Cloud, which host many of these models side by side. That means you can reach several providers through one account and one bill.
Four differences matter more than the rest:
Not all models are the same size
The most useful mental model is a ladder with three rungs. Every provider has models at more than one rung, and the rung matters more than the logo.
Three tiers of model and their relative running cost. Illustrative, not to scale. Compare ratios within one snapshot; prices move constantly.
The gap between rungs is not marginal. Within any single snapshot of prices, a frontier model can cost many times more per task than a small one, often by an order of magnitude or more. Prices move constantly, so the sensible habit is to compare ratios at one point in time rather than remember absolute figures.
Source: Artificial Analysis, artificialanalysis.ai, September 2026. Each model shown at its highest reasoning setting. These figures change all the time, as providers cut prices, release new versions and the benchmark itself is revised, so treat the spread as the point, not the numbers.
Every model in this chart is a capable, current one, yet the most expensive costs around forty times the cheapest on the same set of tasks. Many of them would give an equally acceptable answer to a routine business question. That is the case for choosing deliberately rather than defaulting to whichever name you know.
Where the running cost actually comes from
Almost every provider charges by the token, a unit of text roughly three-quarters of a word long. You pay for tokens going in (your question and any documents attached to it) and for every token the model generates coming out, and output tokens usually cost more than input.
That second part is wider than it sounds. When a model works through a hard problem before answering, it writes out its reasoning as text, usually hidden from you, and every word of that is billed as output. A two-paragraph answer can sit on top of several thousand reasoning tokens. So if a model thinks for a long time, it has almost certainly cost a lot, even though what you see on screen is short.
So the monthly bill is a simple product: how much text goes in, plus how much comes out, multiplied by how often it runs, multiplied by the model's rate. Build cost is paid once. This is paid every month, for as long as the tool is used.
A few habits quietly multiply it
- Pasting the same long document into every request instead of storing it once and retrieving the relevant part.
- Chatty outputs where a one-line answer would do.
- Multi-step automated processes, where one request becomes twenty behind the scenes and the whole conversation so far is re-sent at each step.
- Extended reasoning modes, which are valuable on hard problems and wasteful on easy ones. Many providers let you set how much thinking a model is allowed, and the same question can cost very different amounts depending on that setting.
A worked illustration: a customer-service assistant handling a couple of thousand enquiries a month. Run on a frontier model, the same workload might cost ten or more times what it costs on a mid-tier one, for answers a customer could not tell apart. That difference is decided by a configuration setting, not by how well the tool was built.
Five questions for choosing a model
None of this needs a technical background. It needs the same questions you would ask before buying any piece of equipment.
Choosing a model for a task.
Most business tasks are pattern work dressed up as judgement. Sorting, extracting, summarising and drafting from a template are pattern work. Weighing up a novel situation with conflicting information is judgement. Only the second needs the top rung.
A task that runs ten times a week can afford an expensive model. A task that runs ten thousand times a day cannot, and the difference in quality will not justify it.
Client contracts, staff records and unreleased designs may need to stay inside your own tenancy, which points towards an open model you host rather than a closed one you rent. Public product information can go almost anywhere.
Frontier models are slower. A back-office report that runs overnight does not care. A tool your customer is waiting on does.
Some models handle very long documents in a single pass; others need the material broken up first. If the task is "read the whole tender pack and answer questions", that narrows the field.
Then test. Take twenty real examples from your business, run them through two or three candidate models and compare the answers. It is a half-day exercise, and it is far more reliable than any published benchmark, because it measures the task you actually have.
Don't build your AI twice
The most common mistake we see is not choosing the wrong model. It is building the whole solution inside one vendor's tool, so that the knowledge, the connections to your systems and the rules about how your business does things all live in that tool. When a cheaper or better model arrives, and one will, switching means starting again.
The model is the part that changes. The business knowledge should not have to.
The alternative is to keep those three things somewhere you own, and treat the model as the interchangeable part that plugs into them. Then model choice becomes a monthly decision, made on cost and quality, rather than a one-off commitment you are stuck with.
Where to start
You do not need a strategy for this. You need one task, a clear view of how often it runs and what data it touches, and a short, fair test of the models that could do it. Choose the cheapest that passes, write down why, and look again in a few months.
More insights
Not sure where to begin?
A short conversation about how your business runs and where AI could realistically make a difference. Straight answers either way, including if the answer is not yet.
