Part 1 of 2: The Basics of AI Cost
Daniel Lawrence, CEO, bots for that
Apologies, but yes, another AI buzzword is doing the rounds. Have patience though and stick with me on this one in particular, it’ll be easier but more important to understand than many of the others.
This is the first of two articles. This one is the level-setting part — how AI actually gets costed and charged, in plain terms, so the rest makes sense. Part two takes what’s here and pushes into some more radical territory: what tokenomics actually implies once you stop treating it as a billing curiosity and start treating it as a lens on how (and why) you use AI in the first place.
Let’s start with the title itself, “Tokenomics”. It sounds pretty self-explanatory, unless you don’t know how AI actually incurs cost and is charged, so let’s start there so we don’t overlook the essentials.
Token-who-what?
Tokenomics is the phrase that’s been coined to describe the economics of how AI incurs cost to create and run, and is thereby then charged to users for using it.
There are two main costs associated with AI, both data-centre related, but they work in different ways. Firstly, there’s the cost of storing and processing all the masses of data that models are initially built and trained on. Second, there’s the masses of data involved in our chats — retrieving and generating answers, and holding context and memory while we do it. There’s one other, less direct but not insignificant cost, and that’s the small army of AI engineers, scientists and developers behind all of it — but we’ll park that one for now.
For now, let’s deal with two most important parts:
- how AI actually incurs cost, and how it’s charged on to individuals and businesses
- how this changes the result you can achieve, and why this is important to know
A quick note on the training-versus-running cost split. For most of AI’s recent history, the cost of building and training models was the larger of the two — commonly cited at somewhere around 70% of total AI spend. That’s shifting, and shifting faster than most people realise: several recent industry analyses (Stanford’s AI Index among them) now put the cost of actually running trained models — what’s called inference — at somewhere between 60% and 90% of total AI spend, depending on whose numbers you use and how they’re measuring it. The estimates vary a lot, and I’d treat any single number here with a healthy pinch of salt, but the direction is consistent and worth remembering: training was the big one-off bill, running the thing every day is now the bill that never stops arriving.
This matters practically because venture capital money has funded most of the training cost so far, which has kept user pricing artificially low to encourage adoption. Nobody really knows for sure whether the price we pay will rise sharply as that funding model changes, or whether efficiency gains keep pace and keep prices low anyway. There’s genuine argument on both sides. Only time will tell.
But what exactly is the cost of AI? Surely that’s just the monthly plan fee, right?
In some cases, yes, but not necessarily. There’s a bit more to it than that.
If you’re on a consumer plan (ChatGPT Plus, Claude Pro, Copilot and similar), the plan fee buys you access within a usage limit — typically a cap on the number of messages you can send within a rolling time window, rather than a number you ever see directly. Behind the scenes, the provider is still tracking every token you consume; they’re just absorbing that cost into a flat fee and managing their own exposure by rate-limiting how much you can use. That’s why it feels unlimited even though it isn’t quite.
If instead you’re using AI through an API — connecting a tool, building an automation, running something at scale — there typically is no message cap at all. You pay per token, directly, for exactly what you use. This is where “tokenomics” stops being an abstract concept and starts showing up as a real line on a real invoice.
So, what exactly is a token, and how does it get charged?
Every prompt you submit creates what’s known as a “call” on the AI model and the data centre behind it. That call consumes a variable number of data units, called “tokens”, and every call uses both “input tokens” (roughly, what you typed and any information the AI had to pull in to understand it) and “output tokens” (what it generated back to you). Each is charged separately, and at a noticeably different rate, output tokens typically cost several times more than input tokens, for the straightforward reason that generating new content takes more computational effort than reading existing content.
The exact number of tokens any particular prompt and answer consumes isn’t entirely predictable or consistent, it depends on the model, the language, the formatting, and a fair bit that happens invisibly behind the scenes (system instructions, safety checks, and so on all add tokens you never see).
A worked example
Let’s use your AI tool to draft an email reply.
Imagine you’ve received a 120-word email from a client asking for some information about their account. We’ll assume you’re using an off-the-shelf popular AI tool, connected via an API on a pay-per-use basis (this is where the token cost becomes visible and worth tracking, a consumer subscription would simply absorb this into your flat monthly fee).
Drowning in manual, repetitive work? Tell us the task and we’ll show you what to automate.
The AI generates a 150-word response, acknowledging the query, answering it, and suggesting next steps, all hopefully familiar territory.
As a rough guide, a 120-word input typically translates to somewhere in the region of 150-200 input tokens once you account for the fact that words don’t map one-to-one onto tokens (roughly 1.3 tokens per word in English is a reasonable rule of thumb), plus a modest amount of overhead for system instructions you never see. A 150-word output would consume somewhere around 190-250 output tokens on the same basis. Let’s call it 175 input tokens and 220 output tokens for this example, as middle-of-the-road figures rather than an exact measurement, since as we now know the real number will vary run to run.
At a rate of £0.004 per 1,000 input tokens and £0.02 per 1,000 output tokens, which broadly in line with current flagship-model pricing, where output typically costs somewhere around five to six times as much as input, not just double, because generating new content takes more computational effort than reading existing content:
- Input cost: 175 tokens ÷ 1,000 × £0.004 = £0.0007
- Output cost: 220 tokens ÷ 1,000 × £0.02 = £0.0044
- Total cost per email: roughly £0.005, well under a penny.
If you connected your AI tool to manage a thousand emails a month at that rate, you’d be looking at around £5 a month, which is genuinely small change, and a small fraction of most monthly plan costs.
(These rates are pretty illustrative as of September 2026, based on published flagship-tier API pricing. AI pricing does move fast, and not always in the direction you’d expect, so check your own current provider pricing before budgeting anything at scale).
That’s the straightforward view of it. It does get more complicated and considerably more expensive, once you move into more demanding scenarios, or drive volume and frequency through scale.
For example, I saw someone suggest in a post recently that they’d build their own app using AI. That involves some proper heavy lifting in AI terms: substantial coding, data connections, a front-end interface, back-end architecture, and wiring it all together. That kind of build could easily use several thousand tokens as a one-off development exercise. And once that app is live and running each day on real data for real users, it’ll consume many more tokens on an ongoing basis, every single day. Technically feasible, yes. Commercially viable at scale is a separate question entirely, and one worth actually running the numbers on before you commit.
Five things worth knowing about the cost of AI, that are true for now at least
Token pricing is replacing seat-based SaaS economics. The shift from “per user, per month” to “per token, per call” is the most obvious and most discussed structural change in how AI software is being priced, and it changes the incentives. A seat-based tool costs the same whether you use it once or a hundred times a day. A token-based tool doesn’t.
Cost scales with context, not just output. Longer context windows, retrieved documents, and chat history all get reprocessed on every single call, every time you hit send, the AI is effectively re-reading everything relevant to that conversation, not just your latest message. Three things chiefly drive this: model selection (a more capable model, like the difference between GPT-4 and GPT-5, generally costs more per token); routing decisions (whether the system needs to do a live web search, pull from a knowledge base, or can serve from cached memory, each adds its own cost on top of the base call); and context size (how much background information the AI has to hold in mind to give you a sensible, correct answer).
Agentic workflows multiply cost unpredictably. Loops, retries, and sub-agent delegation mean a single “task” can fan out into dozens of individual calls behind the scenes. This is exactly why a single poorly configured agent, or an unchecked loop, can rack up very significant charges before anyone notices something’s wrong. We’ve seen recent examples of this, for instance, using an AI model to reconcile financial data.
This is also where agentic AI can produce a genuine diseconomy of scale, where the cost per unit of output rises rather than falls as your requirement gets more complex. It’s worth pausing on that, because it not goes against how you expect scale to normally work, but because more and more complex uses are being explored by businesses trying to automate with it. In nature, size usually buys efficiency: an elephant burns dramatically less energy per kilogram of body weight than a mouse does, as bigger, generally, is more efficient per unit. Agentic AI tends to run the opposite way. The more steps, sub-agents and accumulated context a workflow involves, the more the cost per unit of output tends to climb, not fall. More moving parts doesn’t buy you the efficiency you’d instinctively expect, it usually costs you more.
Cost governance is becoming a senior-level control, not an IT line item. The emergence of “FinOps for AI” mirrors the cloud FinOps from a decade ago, with budget caps, per-agent monitoring, and spend alerts, are fast becoming essential, because without them, nobody really notices the runaway cost until the invoice lands.
Right-sizing your models is the obvious lever, but it’s not automatically the right one. Using a smaller, cheaper model for simple tasks, rather than defaulting to the most capable (and most expensive) model available, is the standard cost-control advice, and it’s often good advice. But it’s worth keeping two points in mind: cost optimisation pushes you toward using less capability, while getting genuinely reliable results out of AI is very often about matching the right tool to the task, and using a specialist setup rather than a generalist one. Under-provisioning capability to save money can be a false economy and as much a deployment mistake as over-provisioning it. Cheaper isn’t the same as right. The number of tokens something costs and the quality of judgement it actually requires are two separate questions, and you need to consider both before you decide which model gets the job.
Where this leaves us
That’s the level-setting done. Tokens, calls, input versus output, why a subscription feels unlimited when it isn’t, why an agentic workflow can quietly get expensive, and why the cheapest model isn’t automatically the right model.
None of that is complicated once it’s laid out. Many experts reading this will be waiting for this article to get interesting, and to be honest, I didn’t write this part 1 for you, I wrote it for those business users and decision-makers who either knew nothing or a little but weren’t quite sure. Very few people I’ve talked to who are using AI in a business context could actually explain this clearly. And that’s the point of covering it first. You can’t make a sound decision about how to buy AI, subscription or API, one model or several, a tidy single call or a sprawling agentic workflow, without first laying the foundations and understanding how it’s actually costed underneath the plan you’re looking at. Get this part wrong and every decision built on top of it is guesswork.
The real question tokenomics eventually forces you to ask isn’t “what does this cost?” It’s “what am I actually using AI for, and does the way I’m buying it match that?” A business running a handful of exploratory chats a month has a completely different tokenomics profile, and a completely different “right” way to buy AI, than one running agentic workflows against live customer data all day, every day. Knowing the mechanics is what lets you tell the difference, rather than defaulting to whichever plan looks simplest on a pricing page.
That’s part one. Part two is where it gets more interesting, because once you actually understand how the cost works, some of the more comfortable assumptions about AI adoption start to look a lot less solid.
